BIOIMAGING 2025 Abstracts

Full Papers Short Papers
Full Papers
Paper Nr: 79
Title:

3D View Reconstruction from Endoscopic Videos for Gastrointestinal Tract Surgery Planning
Authors:

Xiaohong W. Gao, Annisa Rahmanti and Barbara Braden
Abstract: This paper investigates the application of neural radiance field (NeRF) to reconstruct a 3D model from 2D endoscopic videos for surgical planning and removal of gastrointestinal lesions. It comprises three stages. The first one is video preprocess to remove frames with artefact of colour misalignment based on a deep learning network. Then the remaining frames are converted into NeRF compatible format. This stage includes extraction of camera information regarding intrinsic, extrinsic and ray pathway parameters as well as conversion to NeRF format based on COLMAP library, a pipeline built upon structure-from-motion (SfM) with multi-view stereo (MVS). Finally the training takes place for establishment of NeRF model implemented upon Nerfstudio library. Initial results illustrate that this end-to-end, i.e. from 2D video input to 3D model output deep learning architecture presents great potentials for reconstruction of gastrointestinal tract. Base on the two sets of data containing 2600 images, the similarity measures of SSIM, PSNR and LPIPS between original (ground truth) and rendered images are 19.46 ± 2.56, 0.70 ± 0.054, and 0.49 ± 0.05 respectively. Future work includes enlarging dataset and removal of ghostly artefact from rendered images.
Download
Paper Nr: 151
Title:

Post-Processing of Thresholding or Deep Learning Methods for Enhanced Tissue Segmentation of Whole-Slide Histopathological Images
Authors:

Michal Marczyk, Agata Wrobel, Julia Merta and Joanna Polanska
Abstract: Digital pathology allows for the efficient storage and advanced computational analysis of stained histopathological slides of various tissues. Tissue segmentation is a crucial first step of digital pathology aimed at eliminating background, pen markings, and other artifacts, reducing image size, and increasing the efficiency of further analysis. In most cases, color thresholding or deep learning models are used, but their effectiveness is reduced due to complex artifacts and huge color variations between slides. We propose a post-processing method to increase the tissue segmentation performance of any initial segmentation algorithm. Using a set of 197 manually annotated histopathological images of breast cancer patients and 63 images of endometrial cancer patients, we tested our method with 3 thresholding techniques and 3 deep learning-based algorithms by calculating the Dice index, Jaccard index, precision, and recall. In both datasets, applying post-processing increased precision and recall for thresholding methods and mostly precision for deep learning models. Overall, applying post-processing gave better tissue segmentation performance than initial segmentation methods, significantly increasing Dice and Jaccard indices. Our results proved that thanks to post-processing, the tissue segmentation pipeline is more robust to noises and artifacts commonly present in histopathological images.
Download
Paper Nr: 227
Title:

Automating Compression Ultrasonography of Human Thigh Tissue and Vessels via Strain Estimation
Authors:

Rytis Jurkonis, Rimvydas Eitminavičius, Vaidotas Marozas and Andrius Sakalauskas
Abstract: Despite the progress made in ultrasonic imaging, the current examination of vein structures by compression is highly operator-dependent and is a time-consuming clinical routine. Current guidelines for the management of deep vein thrombosis recommend compression ultrasonography follow-up for patients at risk of life-threatening complications (pulmonary embolism, heart attack, or stroke). New methods are needed to allow operator-free monitoring of vein structure at the point of care. This article presents the results of integrated imaging with a tissue compression actuator and automated control of tissue deformation through strain estimation. The data for feedback control of the actuator is calculated from raw ultrasound radio-frequency backscattered signals. The region-averaged strain curve (strain versus time) obtained during the tissue compression cycle serves as input for the actuator. The mounting on the human thigh is made from rigid, pre-shaped shells, which are adjusted to the circumference of the thigh with straps. The actuator facilitates a novel, on-body-mounted, automated, operator-free examination of the human femoral vein.
Download
Paper Nr: 244
Title:

Brain MRI Segmentation Using U-Net and SegNet: A Comparative Study Across Modalities with Robust Loss Functions
Authors:

Gouranga Bala, Hiranmay Mondal and Amit Sethi
Abstract: This paper presents a comprehensive comparative study of brain tumor segmentation using two well-known Convolutional Neural Network (CNN) architectures, U-Net and SegNet, across multiple MRI modalities, specifically T2-weighted and Fluid Attenuated Inversion Recovery (FLAIR) images from the BraTS 2020 dataset. We evaluated the performance of these models using four different loss functions: Dice Loss, Focal Loss, Adaptive Robust Loss, and the novel Robust Dice Loss. Our contributions are twofold: first, we provide a detailed comparison of the performance of U-Net and SegNet for brain tumor segmentation across distinct MRI modalities, offering insights into the role of modality-specific features in segmentation outcomes. Second, we introduce the novel Robust Dice Loss, which significantly improved SegNet’s training efficiency, allowing it to handle challenging segmentation scenarios involving data imbalance and intricate tumor boundaries with much greater ease. Our results indicate that U-Net generally outperforms SegNet in terms of segmentation accuracy, particularly when trained with Adaptive Robust Loss. However, the introduction of Robust Dice Loss enabled SegNet to achieve competitive performance, particularly with the FLAIR modality, demonstrating its potential as an effective alternative. This study emphasizes the importance of selecting appropriate loss functions to handle imbalanced data and enhance model performance, thereby contributing valuable insights for the advancement of automated medical image analysis and its clinical utility in neuro-oncology.
Download
Paper Nr: 259
Title:

DEMIS: Electron Microscopy Image Stitching Using Deep Learning Features and Global Optimisation
Authors:

Petr Šilling and Michal Španěl
Abstract: Accurate stitching of overlapping image tiles is essential for reconstructing large-scale Electron Microscopy (EM) images during Whole Slide Imaging. Current stitching approaches rely on handcrafted features and translation-only global alignment based on Minimum Spanning Tree (MST) construction. This results in suboptimal global alignment since it neglects rotational errors and works only with transformations estimated from pairwise feature matches, discarding valuable information tied to individual features. Moreover, handcrafted features may have trouble with repetitive textures. Motivated by the limitations of current methods and recent advancements in deep learning, we propose DEMIS, a novel EM image stitching method. DEMIS uses Local Feature TRansformer (LoFTR) for image matching, and optimises translational and rotational parameters directly at the level of individual features. For evaluation and training, we create EM424, a synthetic dataset generated by splitting high-resolution EM images into arrays of overlapping image tiles. Furthermore, to enable evaluation on unannotated real-world data, we design a no-reference stitching quality metric based on optical flow. Experiments that use the new metric show that DEMIS can improve the average results from 32.11 to 2.28 compared to current stitching techniques (a 1408% improvement). Code is available at: https://github.com/PSilling/demis.
Download
Short Papers
Paper Nr: 40
Title:

Retinoblastoma Detection: Leveraging Deep Learning and Residual Connections for Enhanced Diagnostic Accuracy
Authors:

Shuaa S. Alharbi
Abstract: Retinoblastoma is a rare cancer of the eye that affects children and can be deadly if not diagnosed in time. Detecting this disease early improves the likelihood of curative treatment and makes it possible to preserve the child’s vision. Meanwhile, the application of deep learning techniques to pathology holds the promise of revolutionizing cancer detection and treatment early. When it comes to retinoblastoma, the prospect of automating diagnostic processes to work more accurately and efficiently than healthcare workers can detect dangerous cases with better-than-average accuracy should improve survival rates, as well as rates of vision conservation. In this study, we evaluated several convolutional neural network models: MobileNetV2, EfficientNetB0, ResNet101, DenseNet121, VGG16, and an ensemble model providing a quantitive comparison of which of the models performs best. Among the models, the one that performed best and most accurately was ResNet101, which achieved an accuracy of 97.42%(top-1 accuracy). Comparatively, EfficientNetB0 had a lower metric that indicated its accuracy was 53.40% (top-1 accuracy). ResNet101’s relatively high accuracy for this study suggests that this model is better suited for this type of feature-based classification problem compared to the other models. Residual connection blocks allow layers in a deep neural network to learn to map the input to the same output. This improves performance and reduces errors. Residual networks (ResNets) with many layers have now become the standard architecture used in the leading vision challenges, which gives more insight to researchers and practitioners in choosing the most suitable diagnostic model.
Download
Paper Nr: 47
Title:

Analysis of 3D Urticaceae Pollen Classification Using Deep Learning Models
Authors:

Tijs Konijn, Imaan Bijl, Lu Cao and Fons Verbeek
Abstract: Due to the climate change, hay fever becomes a pressing healthcare problem with an increasing number of affected population, prolonged period of affect and severer symptoms. A precise pollen classification could help monitor the trend of allergic pollen in the air throughout the year and guide preventive strategies launched by municipalities. Most of the pollen classification works use 2D microscopy image or 2D projection derived from 3D image datasets. In this paper, we aim at using whole stack of 3D images for the classification and evaluating the classification performance with different deep learning models. The 3D image dataset used in this paper is from Urticaceae family, particularly the genera Urtica and Parietaria, which are morphologically similar yet differ significantly in allergenic potential. The pre-trained ResNet3D model, using optimal layer selection and extended epochs, achieved the best performance with an F1-score of 98.3%.
Download
Paper Nr: 49
Title:

Assessing the Influence of a CADx Scheme on Radiologists' Analysis of Breast Nodules in Digital Mammography Using Specialized Feedback Software
Authors:

Homero Schiabel, Fernanda J. F. Cardoso and Joyce M. Palotti
Abstract: The study main purpose is to address the effectiveness of a computer-aided diagnosis (CADx) scheme developed to assist radiologists in evaluating nodules in digital mammography images. Unlike traditional CADe systems, which focus primarily on detection, this scheme offers interpretative support, providing additional diagnostic insights for more accurate decisions. This work presents a custom evaluation software designed to facilitate the testing of the CADx scheme influence on radiologists´opinion by allowing them to assess mammograms independently, register their initial opinions, review the CADx output, and log their final decisions. Through this software the study involved radiologists analysing mammograms before and after reviewing the CADx-generated data. The results showed a scheme positive influence on diagnostic accuracy. Radiologists who used the CADx data exhibited in average improved sensitivity and specificity rates, with an overall reduction in error rates, for the images set under investigation. Although the scheme is still a research prototype, it demonstrates strong potential for broader application in clinical practice, offering efficiency and cost-effectiveness, especially for screening operations. The procedure described in this work indicates that, despite the need for some fine-tuning, particularly in minimizing false positives, our CADx system shows promise as a supplemental diagnostic tool that could enhance radiologistśperformance.
Download
Paper Nr: 59
Title:

Enhancing Image Quality to Improve Medical Image Classification: Application to Nuclear Medicine Planar Images
Authors:

Ouassim Boukhennoufa, Laurent Comas, Jean-Marc Nicod, Noureddine Zerhouni and Hatem Boulahdour
Abstract: Nuclear Medicine images are obtained by injecting small amounts of radio-tracers into the body to track specific organs. Particular cameras detect radiations emitted from the radio-tracers resulting in images that visualize the function of the organs rather than their structure. The association of the cameras and radio-tracers causes low resolution and low signal-to-noise ratio, therefore, the images are often of poor quality. Image Quality Enhancement (IQE) is one possible solution to this problem as it improves the clarity of the images by removing noise and correcting distortions. In this paper, we propose a methodology based on artificial intelligence (AI) with the integration of an IQE step for the detection of normal/abnormal parathyroid glands. Two different IQE techniques are employed, one based on a statistical filter and the other on AI. The enhanced images are processed with a Convolutional Neural Network (CNN), and Lasso regression for features extraction and selection. Finally, several AI models are used for binary image classification. The obtained results achieved an accuracy of 83% in distinguishing normal/abnormal parathyroid glands. IQE step significantly improves the accuracy of the AI model by 16.9% over the initial accuracy of 71%, demonstrating the importance of IQE in assessing image classification performance.
Download
Paper Nr: 100
Title:

OCTA Image-Based Machine Learning Models for Discriminating Alzheimer’s Disease from Neurodegenerative and Ocular Conditions
Authors:

Cunyi Xu
Abstract: Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that poses a significant challenge, particularly as the global population ages. Timely diagnosis is crucial for managing AD, and this study aims to contribute to early detection by analyzing Optical Coherence Tomography Angiography (OCTA) images using machine learning models. In this work, we leverage the structural and functional connections between the eye and brain to enhance the discrimination of AD from other neurodegenerative and ocular conditions. We also compiled a comprehensive dataset of OCTA images from various imaging devices, representing a range of diseases. Using a pre-trained nnU-Net, we segmented vascular structures and calculated vascular density metrics, while also extracting histogram and Gray-Level Co-occurrence Matrix (GLCM) features for texture analysis. Various machine learning models were trained and evaluated through five-fold cross-validation, with the Random Forest model achieving 78.15% accuracy in classifying multi-disease OCTA images. The model exhibited high recall for stroke, diabetes, and age-related macular degeneration, but lower recall for AD, congenital heart disease, and hypertension, indicating potential misclassification. Our findings emphasize the utility of OCTA imaging and machine learning for early AD diagnosis, paving the way for future research to refine image processing and classification methods.
Download
Paper Nr: 185
Title:

Bioimages Synthesis and Detection Through Generative Adversarial Network: A Multi-Case Study
Authors:

Valeria Sorgente, Ilenia Verrillo, Mario Cesarelli, Antonella Santone, Fabio Martinelli and Francesco Mercaldo
Abstract: The rapid advancement of Generative Adversarial Networks technology raises ethical and security concerns, emphasizing the need for guidelines and measures to prevent misuse. Strengthening systems to differentiate real from synthetic images and ensuring responsible application in clinical settings could address data scarcity in the biomedical field. For these reasons, considering the increasing popularity of the possibility to generate synthetic images by exploiting artificial intelligence, we investigate the application of Generative Adversarial Networks to generate realistic synthetic bioimages for common pathology representations. We propose a method consisting of two steps: the first one is related to the training of a Deep Convolutional Generative Adversarial Network, while the second step is represented by the evaluation of the bioimages quality using classification-based metrics, comparing synthetic and real images. The model demonstrated promising results, achieving visually realistic images for datasets such as PathMNIST and RetinaMNIST, with accuracy improving over training epochs. However, challenges arose with datasets like ChestMNIST and OCTMNIST, where image quality was limited, showing poor detail and distinguishability from real samples.
Download
Paper Nr: 220
Title:

Pilot Study of Distinct Graphs Models in Analysis of Brain Aging in Resting-State Functional Connectivity Networks
Authors:

M. A. G. Carvalho and R. Frayne
Abstract: Graphs have been used successfully to represent and analyze brain networks for many decades. Such structural and functional studies are important for revealing interactions between distinct areas of the brain that, for example, are associated with the performance of a specific task or the onset of a cognitive disorder like dementia. In this pilot study, resting-state functional magnetic resonance imaging data were acquired in a sex-balanced sample of 10 young (20.1±2.1 years) and 10 old (65.6±0.4 years), presumed healthy, adults. We examined the effects of age on whole-brain resting-state functional connectivity (RSFC) networks. We examined two main graph modeling approaches to analyze RSFC networks. These approaches employ different strategies or graph models for thresholding over the complete network or examining changes in graph density. We computed and compared one graph metric, the modularity, that was derived from the RSFC network graph models. Considering the need for a model that must preserve the network’s connectivity, strategies that use spanning trees as seeds to gradually increase the graph’s density seem more appropriate to represent brain networks.
Download
Paper Nr: 221
Title:

First Results on Graph Similarity Search in Resting-State Functional Connectivity Networks Using Spectral and Graph Edit Distances
Authors:

M. A. G. Carvalho and R. Frayne
Abstract: The application of graph theory in the modeling and analysis of brain networks has generated both new opportunities as well as new challenges in neuroscience. Resting state functional connectivity (RSFC) networks studied with graphs is an important field of investigation because of the potential benefits in understanding function in healthy individuals and identifying evidence of brain diseases and injury in patients. This work is unique because it applies information retrieval techniques to create ranked lists from RSFC graph theory-derived networks. In our analysis, we used a sample of whole-brain resting-state functional magnetic resonance imaging (rs-fMRI) data obtained from Young (n = 10, age: 20.1 ± 2.1) and Old (n = 10, 65.6 ± 0.4) sex-balanced groups drawn from a healthy, i.e., neurotypical, cohort. We estimated two well-known distance metrics (graph edit distance and graph spectral distance) and by using information-retrieval graph ranking methods achieved precision measures at the top-5 positions of ranked lists of up to 80%.
Download
Paper Nr: 226
Title:

Preliminary Results on Using Clustering of Functional Data to Identify Patients with Alzheimer’s Disease by Analyzing Brain MRI Scans
Authors:

Calin Anton, Cristina Anton, Mohamad El-Hajj, Matthew Craner and Richard Lui
Abstract: This study delves into the effectiveness of funWeightClust, a sophisticated model-based clustering technique that leverages functional linear regression models to pinpoint patients diagnosed with Alzheimer’s Disease. Our research entailed a thorough analysis of voxelwise fractional anisotropy data derived from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, with a particular emphasis on the Cingulum and Corpus Callosum, which are critical regions of interest in understanding the disease’s impact on brain structure. Through a series of experiments, we established that funWeightClust is efficient at distinguishing between patients with Alzheimer’s Disease and healthy control subjects. Notably, the clustering model yielded even more pronounced and accurate results when we focused our analysis on specific brain regions, such as the Left Hippocampus and the Splenium. We postulate that integrating additional biomarkers could significantly enhance the accuracy and reliability of funWeightClust in identifying patients who exhibit signs of Alzheimer’s Disease.
Download
Paper Nr: 241
Title:

Investigating Reinforcement Learning for Histopathological Image Analysis
Authors:

Mohamad Mohamad, Francesco Ponzio, Maxime Gassier, Nicolas Pote, Damien Ambrosetti and Xavier Descombes
Abstract: In computational pathology, whole slide images represent the primary data source for AI-driven diagnostic algorithms. However, due to their high resolution and large size, these images undergo a patching phase. In this paper, we approach the diagnostic process from a pathologist’s perspective, modeling it as a Sequential decision-making problem using reinforcement learning. We build a foundational environment designed to support a range of whole slide applications. We showcase its capability by using it to construct a toy goal-conditioned Navigation environment. Finally, we present an agent trained within this environment and provide results that emphasize both the promise of reinforcement learning in histopathology and the distinct challenges it faces.
Download
Paper Nr: 248
Title:

Deep Learning Denoising of Low-Dose Computed Tomography Using Convolutional Autoencoder: A Phantom Study
Authors:

Simone Damiani, Manuela Imbriani, Francesca Lizzi, Mariagrazia Quattrocchi, Alessandra Retico, Sara Saponaro, Camilla Scapicchio, Alessandro Tofani, Arman Zafaranchi, Maria Irene Tenerani and Maria Evelina Fantacci
Abstract: Low Dose Computed Tomography (LDCT) has proven to be an optimal clinical exploration method for early diagnosis of lung cancer in asymptomatic but high-risk population; however, this methodology suffers from a considerable increase in image noise with respect to Standard Dose Computed Tomography (CT) scans. Several approaches, both conventional and Deep Learning (DL) based, have been developed to mitigate this problem while preserving the visibility of the radiological signs of pathology. This study aims to exploit the possibility of using DL-based methods for the denoising of LDCTs, using a Convolutional Autoencoder and a paired low-dose and high-dose scans (LD/HD) dataset of phantom images. We used twelve acquisitions of the Catphan-500® phantom, each containing 130 slices, acquired with two CT scanners, two dose levels and reconstructed using the Filtered BackProjection algorithm. The proposed architecture, trained with a com-bined loss function, shows promising results for both noise magnitude reduction and Contrast-to-Noise Ratio enhancement when compared with HD reference images. These preliminary results, while encouraging, leave a wide margin for improvement and need to be replicated first on phantoms with more complex structures, secondly on images reconstructed with Iterative Reconstruction algorithms and then translated to LDCTs of real patients.
Download
Paper Nr: 279
Title:

Deep Learning-Based Classification of Stress in Sows Using Facial Images
Authors:

Syed U. Yunas, Ajmal Shahbaz, Emma M. Baxter, Marianne Farish, Kenneth M. D. Rutherford, Mark F. Hansen, Melvyn L. Smith and Lyndon N. Smith
Abstract: Stress in pigs is a significant factor contributing to poor health, increased antimicrobial usage, and the subsequent risk of antimicrobial resistance (AMR), which poses a major challenge for the global pig farming industry. In this paper, we propose using deep learning (DL) methods to classify stress levels in sows based on facial features captured from images. Early identification of stress can enable targeted interventions, potentially reducing health risks and mitigating AMR concerns. Our approach utilizes convolutional neural network (CNN) models, specifically YOLO8l-cls, to classify the stress levels of sows (pregnant pigs) into low-stressed and high-stressed categories. Experimental results demonstrate that YOLO8l-cls outperforms other classification methods, with an overall F1-score of 0.74, Cohen’s Kappa of 0.63, and MCC of 0.60. This highlights the model’s effectiveness in accurately identifying stress levels and its potential as a practical tool for stress management in pig farming, with benefits for animal welfare, the farming industry, and broader efforts to minimize AMR risk.
Download
Paper Nr: 316
Title:

Use of Radiomics in Low Dose Chest CT: A Proposal for a Phantom Multi-Centric Study
Authors:

Maria Irene Tenerani, Silvia Arezzini, Antonino Formuso, Francesca Lizzi, Enrico Mazzoni, Stefania Pallotta, Alessandra Retico, Camilla Scapicchio, Cinzia Talamonti and Maria Evelina Fantacci
Abstract: Radiomics is a quantitative biomedical image analysis tool involving the mathematical extraction of image features that can be used, particularly in oncology, to build predictive models based on artificial intelligence for diagnosis and treatment outcome prediction. In Lung cancer screening via Low-Dose Computed Tomography (LDCT), radiomics-based models could increase lung nodules detectability simplifying the implementation of large-scale screening. However, their transposition into clinical practice is slowed by the instability that radiomic feature values show in changes in CT image acquisition and reconstruction parameters. To build more robust models, it is essential to conduct multi-centric radiomic studies leveraging the use of various types of phantoms to overcome the challenges associated with patient data complexity. However, many difficulties may arise related to both the image acquisition and reconstruction process and the extraction and analysis of radiomic features. In this paper, from the results of a pilot study conducted with two phantoms, guidelines for a multi-centric radiomic study on phantoms LDCTs are proposed, focusing on crucial aspects such as phantom positioning, image acquisition and reconstruction protocol, and radiomic feature extraction pipeline. Finally, a XNAT-based platform for data sharing and management, image quality control implementation and radiomic feature extraction automation is proposed.
Download
Paper Nr: 326
Title:

3D Visualization and Interaction for Studying Respiratory Infections by Exploiting 2D CNN-Derived Attention Maps and Lung Models
Authors:

Mohamed El Fateh Hadjarsi, Adnan Mustafic, Mahmoud Melkemi, Iyed Dhahri and Karim Hammoudi
Abstract: Nowadays, research activities in the fields of precision health and biomedical image analysis are developing rapidly. In this context, research work on the analysis of respiratory infections is still extensively investigated. Few open source systems with the goal of visualizing and manipulating lungs with infections in 3D space are currently proposed. Such systems could become an important tool in the training of new radiologists. In the present work, we propose an approach that allows the user to visualize and interact with respiratory infections in 3D space by exploiting 2D CNN-derived attention maps. The source code will be made publicly available at https://github.com/Adn-an/3D-Visualization-and-Interaction-for-Studying-Respiratory-Infections-by-Exp loiting-2D-Attention-Maps.
Download
Paper Nr: 43
Title:

Using MIRNet for Low Light Image Enhancement
Authors:

Ethan Chen, Robail Yasrab and Pramit Saha
Abstract: This study explores the application of MIRNet (Multi-scale Image Restoration Network), a deep learning architecture designed for image enhancement. MIRNet uses convolutional neural networks (CNNs) to capture image details and textures at various scales, enabling effective restoration and enhancement of low-quality images. Experiments were conducted using the LoL and SICE datasets to validate and optimize MIRNet’s performance. The results were compared with an existing image enhancement tool, demonstrating the superior effectiveness of MIRNet even with architectural modifications or training on different data sets. The research also explains MIRNet’s architecture and its approach to processing and enhancing image content. This work highlights MIRNet’s potential to advance image enhancement through deep learning techniques.
Download
Paper Nr: 77
Title:

Using Machine Learning to Identify Crop Diseases with ResNet-18
Authors:

Rihan Rahman
Abstract: Plant diseases are a highly prevalent issue in agriculture, causing countless farmers annually to face career threatening damages such as diminished profits and crop yields and environmental damages. Consequently, it is imperative that these diseases are quickly detected and treated against. An increasingly effective solution is to train convolutional neural networks (CNNs) using deep learning (DL). DL has several effective applications in a variety of major fields such as healthcare and fraud detection and has a high potential to solve issues of global significance. This research’s goal is to create a machine learning (ML) model with DL to identify plants’ diseases using photos of infected leaves. Many farmers in rural areas struggle to treat blights due to limited access to technology and information regarding them. Therefore, an ML model which can automatically identify these diseases would be highly useful for these people. After sourcing a comprehensive dataset with images of 88 types of plants and diseases, I used it to train a CNN model using several data augmentation techniques. With the model architecture ResNet-18, while evaluating its performance with a validation dataset, the model achieved a loss of 4.541%. This value demonstrates ResNet-18’s applicability to the task of identifying plant diseases and illustrates the potential for classification-based DL networks to support rural farmers and the field of agriculture. If a superior model is created to identify blights more accurately, it should be used to help the billions of farmers who would greatly benefit from such technology.
Download
Paper Nr: 87
Title:

A Real-World Segmentation Model for Melanocytic and Nonmelanocytic Dermoscopic Images
Authors:

Eleonora Melissa, Daria Riabitch, Linda Lazzeri, Federica La Rosa, Chiara Benvenuti, Mario D’Acunto, Giovanni Bagnoni, Daniela Massi and Marco Laurino
Abstract: Segmentation is a critical step in computer-aided diagnosis (CAD) systems for skin lesion classification. In this study, we applied the Deeplabv3+ network to segment real dermoscopic images. The model was trained on public datasets and tested both on public and on a disjoint set of images from the TELEMO project, covering six clinically significant skin lesion types: basal cell carcinoma, squamous cell carcinoma, melanoma, benign nevi, actinic keratosis and seborrheic keratosis. Our model achieved a testing global accuracy of 97.88% on public dataset and of 92.62% on TELEMO dataset, outperforming literature models. Although some misclassifications occurred, largely due to class imbalance, the model demonstrated strong generalization to real-world clinical images, a critical achievement for deep learning in medical imaging. To evaluate the clinical relevance of our segmentation, we extracted ten key features related to lesion border and diameter. Notably, the ”Diameters Mean” and ”Area to Perimeter Product” features revealed significant differences between melanoma-nevi and basal cell carcinoma-nevi classes, with strong effect sizes. These findings suggest that border features are crucial for distinguishing between multiple skin lesion types, highlighting the model’s potential for aiding dermatological diagnoses.
Download
Paper Nr: 198
Title:

Evaluation of OCT Image Synthesis for Choroidal and Retinal Layer Segmentation Using Denoising Diffusion Probabilistic Models
Authors:

Yudai Yamauchi, Yuli Wu and Eiji Okada
Abstract: Machine learning can automatically conduct the layer segmentation task of retinal optical coherence tomography (OCT) image, but annotated data is required to train these models. Synthetic retinal OCT images are generated using denoising diffusion probabilistic models (DDPMs), which can be used to train segmentation models effectively and automatically create annotated data. However, the extent to which these synthetic images contribute to segmentation accuracy compared to real data has not been investigated. In this study, we synthesized retinal OCT images from sketch images using DDPMs, trained a segmentation model using synthetic and real images, and evaluated how the use of synthetic images influenced the accuracy of choroidal and retinal layer segmentation compared to results using only real images. Through a comparison of the Dice score, we confirmed that training with both synthetic and real OCT images led to higher Dice scores than training with only real OCT images. These findings suggest that using synthetic images can enhance segmentation accuracy, offering a promising approach to improving model performance in situations with limited annotated real data.
Download
Paper Nr: 275
Title:

U-Net in Histological Segmentation: Comparison of the Effect of Using Different Color Spaces and Final Activation Functions
Authors:

László Körmöczi and László G. Nyúl
Abstract: Deep neural networks became widespread in numerous fields of image processing, including semantic segmentation. U-Net is a popular choice for semantic segmentation of microscopy images, e.g. histological sections. In this paper, we compare the performance of a U-Net architecture in three different color spaces: the commonly used, perceptually uniform sRGB, the perceptually uniform but device-independent CIE L*a*b*, and linear RGB color space that is uniform in terms of light intensity. Furthermore, we investigate the network’s performance on data combinations that were unseen during training.
Download
Paper Nr: 305
Title:

Stratum Corneum Light Confinement: Monte Carlo Verification
Authors:

Leah DeVos, Gennadi Saiko and Alexandre Douplik
Abstract: Significance: The epidermis, the outermost layer of the skin, plays a crucial role in protecting the body from UV radiation, chemical substances, and physical trauma. Its top layer, the stratum corneum (SC), consists of dead skin cells with low water content (~20%), creating a refractive index gradient between the SC and underlying tissue. This gradient traps light within the SC layer, but its impact on light propagation in tissues remains largely unexplored. Aim: The study investigates how refractive index variations in the skin influence light propagation in tissues. Approach: Monte Carlo (MC) light transport simulations were performed in media with and without refractive index mismatches. Results: Light confinement in the SC increases the fluence rate by 12-35% compared to underlying tissue, particularly when the underlying tissue has low diffuse reflectance. This effect is most pronounced when the SC thickness exceeds the reduced scattering length (~150-600 μm for visible light). Such thicknesses occur in glabrous skin (palms, soles) and thickened areas like calluses and corns. Conclusions: By comparing MC simulations, we attribute this light confinement to the SC's high refractive index due to its low water content. This stratum corneum light confinement (SCLC) phenomenon may lead to an inaccurate estimation of light distribution, resulting in errors in some skin diagnostic parameters measured via diffuse reflection, such as water and total hemoglobin content, and blood oxygenation.
Download