Convolutional Neural Network Training Research Articles

Convolutional Neural Networks Assisted Peak Classification in Targeted LC-HRMS/MS for Equine Doping Control Screening Analyses.

Doping control screening analyses usually involve visual inspection of extracted ion chromatograms (EIC) by a trained analytical chemist, followed by further investigations if needed. This task is both highly repetitive and time-consuming, given the hundreds of compounds and metabolites to be screened in tens of thousands of samples per year. With the recent widespread adoption of machine learning in analytical chemistry and the training of high-performance convolutional neural networks (CNN), these operations can be automated with high accuracy and throughput. Applying this technology to doping control is challenging as the false negative rate (FNR) shall be equal to zero. In this study, we demonstrated that implementing a deep learning strategy for chromatogram classification in equine doping control can be feasible and accurate. We illustrated our findings with a CNN scoring model combined with a linear discriminant analysis (LDA) classifier trained on chromatogram images from our ultra-high-pressure liquid chromatography coupled to high-resolution tandem mass spectrometry (UHPLC-HRMS/MS)-based biotherapeutics screening method. We expect that artificial intelligence (AI) will be a valuable tool for doping control laboratories in the near future.

Convolutional Neural Network for Depression and Schizophrenia Detection

Background/Objectives: This study presents a Convolutional Neural Network (CNN) approach for detecting depression and schizophrenia using motor activity patterns represented as images. Participants’ motor activity data were captured and transformed into visual representations, enabling advanced computer vision techniques for the classification of these mental disorders. The model’s performance was evaluated using a three-fold cross-validation, achieving an average accuracy of 95%, demonstrating the effectiveness of the proposed approach in accurately identifying mental health conditions. The objective of the study is to develop a model capable of identifying different mental disorders by processing motor data using CNN in order to provide a support tool to areas specialized in the diagnosis and efficient treatment of these psychological conditions. Methods: The methodology involved segmenting and transforming motor activity data into images, followed by a CNN training and testing phase on these visual representations. This innovative approach enables the identification of subtle motor behavior patterns, potentially indicative of specific mental states, without invasive interventions or self-reporting. Results: The results suggest that CNNs can capture discriminative features in motor activity to differentiate between individuals with depression, schizophrenia, and those without mental health diagnoses. Conclusions: These findings underscore the potential of computer vision and deep neural network techniques to contribute to early, non-invasive mental health disorder diagnosis, with significant implications for developing mental health support tools.

Application of convolutional neural networks for mobile internet forecasting in Colombia

The growth of mobile Internet has facilitated access to information by minimizing geographical barriers. For this reason, this paper forecasts the number of users, incomes, and traffic for operators with the most significant penetration in the mobile internet market in Colombia to analyze their market growth. For the forecast, the convolutional neural network (CNN) technique is used, combined with the recurrent neural network (RNN), long short-term memory network (LSTM), and gated recurrent unit (GRU) techniques. The CNN training data corresponds to the last twelve years. The results currently show a high concentration in the market since a company has a large part of the market; however, the forecasts show a decrease in its users and revenues and the growth of part of the competition. It is also concluded that the technique with the most precision in the forecasts is CNN-GRU.

Representative training data sets are critical for accurate machine-learning classification of microscopy images of particles formed by lipase-catalyzed polysorbate hydrolysis.

Polysorbate 20 (PS20) is commonly used as an excipient in therapeutic protein formulations. However, over the course of a therapeutic protein product's shelf life, minute amounts of co-purified host-cell lipases may cause slow hydrolysis of PS20, releasing fatty acids (FAs). These FAs may precipitate to form subvisible particles that can be detected and imaged by various techniques, e.g., flow imaging microscopy (FIM). Images of particles can then be classified using supervised convolutional neural networks (CNNs). However, CNNs should be trained on representative images of particles which, as we demonstrate in this work, may be challenging to obtain. Here, we tested several rapid techniques to create FA particles and examined whether CNNs trained on microscopy images of these rapidly formed particles could accurately classify images of particles that had been produced by kinetically slower lipase-catalyzed hydrolysis of PS20. CNNs trained on images of rapidly produced particles were less accurate in classifying images of FA particles that had been produced by enzymatic hydrolysis of PS20 than CNNs trained with images of particles generated by the same slow hydrolysis, highlighting the importance of using representative image data sets for training CNN classifiers.

Advancing explainability of adversarial trained Convolutional Neural Networks for robust engineering applications

The great success of Convolutional Neural Networks has driven their widespread use in Industry 5.0, especially in safety-critical applications where model robustness is essential. While adversarial training is commonly used to make models more robust, it often changes how the model fuses features, leading to complex trade-offs and making its behavior harder to understand. This lack of transparency contradicts the demand for interpretable and trustworthy artificial intelligence systems in modern industry, where applications like autonomous driving and intelligent medical systems require model interpretability to assign post-accident responsibility and support legal and liability decisions. Our work introduces an innovative approach to enhance the interpretability of adversarially trained Convolutional Neural Networks, providing detailed and quantifiable insights through a six-dimensional semantic framework. This framework, which examines objects, parts, scenes, materials, textures, and colors, offers fine-grained explanations that closely mirror human cognitive processes. We conduct extensive experiments across multiple classical models, and the results reveal that adversarial training induces structure-dependent changes in feature preferences, offering insights into the trade-off between standard and robust accuracy in adversarially trained Convolutional Neural Networks. These findings provide rational support for the use of artificial intelligence in real-world applications and promote the development and implementation of explainable artificial intelligence in modern industries.

Open-source Large Language Models can Generate Labels from Radiology Reports for Training Convolutional Neural Networks

Open-source Large Language Models can Generate Labels from Radiology Reports for Training Convolutional Neural Networks

Qualify-as-you-go: sensor fusion of optical and acoustic signatures with contrastive deep learning for multi-material composition monitoring in laser powder bed fusion process

ABSTRACT Growing demand for multi-material Laser Powder Bed Fusion (LPBF) faces process control and quality monitoring challenges, particularly in ensuring precise material composition. This study explores optical and acoustic emission signals during LPBF processes with multiple materials, addressing challenges in process control and ensuring accurate material composition. Experimental data from processing five powder compositions were collected using a custom-built monitoring system in a commercial LPBF machine. The research categorised signals from LPBF processing various compositions, enhancing prediction accuracy by combining optical with acoustic data and training convolutional neural networks using contrastive learning. Latent spaces of trained models using two contrastive loss functions, clustered acoustic and optical emissions based on similarities, aligning with five compositions. Contrastive learning and sensor fusion were found to be essential for monitoring LPBF processes involving multiple materials. This research advances the understanding of multi-material LPBF, highlighting sensor fusion strategies’ potential for improving quality control in additive manufacturing.

Leveraging domain knowledge for synthetic ultrasound image generation: a novel approach to rare disease AI detection.

This study explores the use of deep generative models to create synthetic ultrasound images for the detection of hemarthrosis in hemophilia patients. Addressing the challenge of sparse datasets in rare disease diagnostics, the study aims to enhance AI model robustness and accuracy through the integration of domain knowledge into the synthetic image generation process. The study employed two ultrasound datasets: a base dataset (Db) of knee recess distension images from non-hemophiliac patients and a target dataset (Dt) of hemarthrosis images from hemophiliac patients. The synthetic generation framework included a content generator (Gc) trained on Db and a context generator (Gs) to adapt these images to match Dt's context. This approach generated a synthetic target dataset (Ds), primed for AI training in rare disease research. The assessment of synthetic image generation involved expert evaluations, statistical analysis, and the use of domain-invariant perceptual distance and Fréchet inception distance for quality measurement. Expert evaluation revealed that images produced by our synthetic generation framework were comparable to real ones, with no significant difference in overall quality or anatomical accuracy. Additionally, the use of synthetic data in training convolutional neural networks demonstrated robustness in detecting hemarthrosis, especially with limited sample sizes. This study presents a novel approach for generating synthetic ultrasound images for rare disease detection, such as hemarthrosis in hemophiliac knees. By leveraging deep generative models and integrating domain knowledge, the proposed framework successfully addresses the limitations of sparse datasets and enhances AI model training and robustness. The synthetic images produced are of high quality and contribute significantly to AI-driven diagnostics in rare diseases, highlighting the potential of synthetic data in medical imaging.

Design and application of CNN for emission detection through thermal imagery

Motorcycle exhaust emissions (EE) that do not meet regulatory standards present a significant environmental and public health issue, particularly given the rising number of motorcycles in densely populated areas. These emissions release pollutants such as carbon monoxide (CO), hydrocarbons (HC), and nitrogen oxides (NOx), which contribute to poor air quality and have adverse effects on human health. Traditional emission testing methods using gas analyzers, while commonly used, face limitations such as sensitivity to environmental fluctuations, the necessity for frequent recalibration, and an intensive testing process requiring specialized expertise. This study addresses these issues by developing an innovative method for emission detection using Convolutional Neural Networks (CNN) applied to thermal images of motorcycle exhausts. The research method involves five key stages: data acquisition, dataset formation, CNN model design and training, model testing, and validation. Thermal images were gathered from 27 motorcycles, representing various brands and engine configurations common in Indonesia, and each image set included 100 samples for both emission-compliant and non-compliant categories. By analyzing thermal patterns, the CNN model was trained to accurately detect combustion patterns indicative of emission status based on the lambda value. This approach enables the model to generalize across different motorcycle models, offering practical adaptability for widespread implementation. The results demonstrate that the CNN model delivers high predictive accuracy, precision, recall, and F1-score, making it a robust tool for assessing motorcycle emission compliance. This CNN-based approach provides a practical solution for real-time, large-scale emission monitoring and regulatory enforcement, reducing dependency on conventional methods. Its scalability and adaptability position it as a valuable advancement in emission monitoring technology, with significant potential for supporting environmental standards and improving air quality management

AneRBC dataset: a benchmark dataset for computer-aided anemia diagnosis using RBC images.

Visual analysis of peripheral blood smear slides using medical image analysis is required to diagnose red blood cell (RBC) morphological deformities caused by anemia. The absence of a complete anaemic RBC dataset has hindered the training and testing of deep convolutional neural networks (CNNs) for computer-aided analysis of RBC morphology. We introduce a benchmark RBC image dataset named Anemic RBC (AneRBC) to overcome this problem. This dataset is divided into two versions: AneRBC-I and AneRBC-II. AneRBC-I contains 1000 microscopic images, including 500 healthy and 500 anaemic images with 1224 × 960 pixel resolution, along with manually generated ground truth of each image. Each image contains approximately 1550 RBC elements, including normocytes, microcytes, macrocytes, elliptocytes, and target cells, resulting in a total of approximately 1 550 000 RBC elements. The dataset also includes each image's complete blood count and morphology reports to validate the CNN model results with clinical data. Under the supervision of a team of expert pathologists, the annotation, labeling, and ground truth for each image were generated. Due to the high resolution, each image was divided into 12 subimages with ground truth and incorporated into AneRBC-II. AneRBC-II comprises a total of 12 000 images, comprising 6000 original and 6000 anaemic RBC images. Four state-of-the-art CNN models were applied for segmentation and classification to validate the proposed dataset. Database URL: https://data.mendeley.com/preview/hms3sjzt7f?a=4d0ba42a-cc6f-4777-adc4-2552e80db22b.

A Combined CNN-LSTM Network for Ship Classification on SAR Images.

Satellite SAR (synthetic aperture radar) imagery offers global coverage and all-weather recording capabilities, making it valuable for applications like remote sensing and maritime surveillance. However, its use in machine learning-based automatic target classification faces challenges, including the limited availability of SAR target training samples and the inherent constraints of SAR images, which provide less detailed features compared to natural images. These issues hinder the effective training of convolutional neural networks (CNNs) and complicate the transfer learning process due to the distinct imaging mechanisms of SAR and natural images. To address these challenges, we propose a shallow CNN architecture specifically designed to optimize performance on SAR datasets. Evaluations were performed on three datasets: FUSAR-Ship, OpenSARShip, and MSTAR. While the FUSAR-Ship and OpenSARShip datasets present difficulties due to their limited and imbalanced class distributions, MSTAR serves as a benchmark with balanced classes. To compare and optimize the proposed shallow architecture, we examine various properties of CNN components, such as the filter numbers and sizes in the convolution layers, to reduce redundancy, improve discrimination capability, and decrease network size and learning time. In the second phase of this paper, we combine the CNN with Long short-term memory (LSTM) networks to enhance SAR image classification. Comparative experiments with six state-of-the-art CNN architectures (VGG16, ResNet50, Xception, DenseNet121, EfficientNetB0, and MobileNetV2) demonstrate the superiority of the proposed approach, achieving competitive accuracy while significantly reducing training times and network complexity. This study underscores the potential of customized architectures to address SAR-specific challenges and enhance the efficiency of target classification.

Automating tephra fall building damage assessment using deep learning

Abstract. In the wake of a volcanic eruption, the rapid assessment of building damage is paramount for effective response and recovery planning. Uncrewed aerial vehicles, UAVs, offer a unique opportunity for assessing damage after a volcanic eruption, with the ability to collect on-demand imagery safely and rapidly from multiple perspectives at high resolutions. In this work, we established a UAV-appropriate tephra fall building damage state framework and used it to label ∼50 000 building bounding boxes around ∼2000 individual buildings in 2811 optical images collected during surveys conducted after the 2021 eruption of La Soufrière volcano, St Vincent and the Grenadines. We used these labelled data to train convolutional neural networks (CNNs) for (1) building localisation (average precision equals 0.728) and (2) damage classification into two levels of granularity: no damage vs. damage (F1 score = 0.809) and moderate damage vs. major damage (F1 score = 0.838) (1 is the maximum obtainable for both metrics). The trained models were incorporated into a pipeline along with all the necessary image processing steps to generate spatial data (a georeferenced vector with damage state attributes) for rapid tephra fall building damage mapping. Using our pipeline, we assessed tephra fall building damage for the town of Owia, finding that 22 % of buildings that received 50–90 mm of tephra accumulation experienced at least moderate damage. The pipeline is expected to perform well across other volcanic islands in the Caribbean where building types are similar, though it would benefit from additional testing. Through cross-validation, we found that the UAV look angle had a minor effect on the performance of damage classification models, while for the building localisation model, the performance was affected by both the look angle and the size of the buildings in images. These observations were used to develop a set of recommendations for data collection during future UAV tephra fall building damage surveys. This is the first attempt to automate tephra fall building damage assessment solely using post-event data. We expect that incorporating additional training data from future eruptions will further refine our model and improve its applicability worldwide. To facilitate continued development and collaboration all trained models and the pipeline code can be downloaded from GitHub.

Optimising Concrete Crack Detection: A Study of Transfer Learning with Application on Nvidia Jetson Nano.

The use of Artificial Intelligence (AI) to detect defects such as concrete cracks in civil and transport infrastructure has the potential to make inspections less expensive, quicker, safer and more objective by reducing the need for on-site human labour. One deployment scenario involves using a drone to carry an embedded device and camera, with the device making localised predictions at the edge about the existence of defects using a trained convolutional neural network (CNN) for image classification. In this paper, we trained six CNNs, namely Resnet18, Resnet50, GoogLeNet, MobileNetV2, MobileNetV3-Small and MobileNetV3-Large, using transfer learning technology to classify images of concrete structures as containing a crack or not. To enhance the model's robustness, the original dataset, comprising 3000 images of concrete structures, was augmented using salt and pepper noise, as well as motion blur, separately. The results show that Resnet50 generally provides the highest validation accuracy (96% with the original dataset and a batch size of 16) and the highest validation F1-score (95% with the original dataset and a batch size of 16). The trained model was then deployed on an Nvidia Jetson Nano device for real-time inference, demonstrating its capability to accurately detect cracks in both laboratory and field settings. This study highlights the potential of using transfer learning on Edge AI devices for Structural Health Monitoring, providing a cost-effective and efficient solution for automated crack detection in concrete structures.

An explainable framework for convolutional neural networks detecting dementia in MRI scans

AbstractBackgroundWith a global ageing population, there is an increasing demand for fast and reliable early diagnosis of individuals. Convolutional neural networks (CNNs) have an immense potential in assisting clinicians in diagnosing dementia. Regional atrophy patterns, which are visible in T1‐weighted MRI scans, have been consistently identified by the CNNs with high accuracy. However, there is still hesitation in the clinical adoption of CNNs, due to their black box nature. To alleviate these issues, we propose a framework for explainable CNN models to detect neurodegenerative diseases.MethodUsing data from multiple data cohorts (N>3000) we trained a robust DenseNet CNN model and utilised the feature attribution method ‘layer‐wise relevance propagation’ (LRP) to highlight image features found useful by the model. The FastSurfer tool was utilised to segment individual brain scans into various anatomical region of interests (ROIs). We created an ontology model for encoding the hierarchy of and relationships between neuroanatomical structures. Using linear regression models, controlling for common covariates such as age and sex, we quantified the volumetric deviance from normal levels (w‐scores) for all ROIs. These findings were visualised as summary plots. We integrate two different sources of information about pathology in the MRI scans, 1) the CNN predictive models and 2) the volumetric atrophy analysis. An algorithmic framework was developed to combine both of them and to select the information to be included in textual explanations for the CNN’s predictions (Figure 1).ResultOur framework results in a modular pipeline where each component – attribution method, segmentation tool, ontology model, could be altered due to advancements or changes in the requirements. Our framework implements a plug‐and‐play functionality for interpreting trained CNN models. It assists an easy understanding of CNN’s prediction by generating textual and visual explanations for its findings, with variable degree of detail, focusing on specific ROIs or lobes.ConclusionOur study provides a proof‐of‐concept of utilizing a‐priori neuroanatomical information for explaining CNNs. This prototypical implementation for generating explainable AI is going to be tested by clinicians to evaluate its utility for assisting the diagnosis of dementia. We also plan to extend our framework to other neurodegenerative diseases.

An explainable framework for convolutional neural networks detecting dementia in MRI scans

AbstractBackgroundWith a global ageing population, there is an increasing demand for fast and reliable early diagnosis of individuals. Convolutional neural networks (CNNs) have an immense potential in assisting clinicians in diagnosing dementia. Regional atrophy patterns, which are visible in T1‐weighted MRI scans, have been consistently identified by the CNNs with high accuracy. However, there is still hesitation in the clinical adoption of CNNs, due to their black box nature. To alleviate these issues, we propose a framework for explainable CNN models to detect neurodegenerative diseases.MethodUsing data from multiple data cohorts (N>3000) we trained a robust DenseNet CNN model and utilised the feature attribution method ‘layer‐wise relevance propagation’ (LRP) to highlight image features found useful by the model. The FastSurfer tool was utilised to segment individual brain scans into various anatomical region of interests (ROIs). We created an ontology model for encoding the hierarchy of and relationships between neuroanatomical structures. Using linear regression models, controlling for common covariates such as age and sex, we quantified the volumetric deviance from normal levels (w‐scores) for all ROIs. These findings were visualised as summary plots. We integrate two different sources of information about pathology in the MRI scans, 1) the CNN predictive models and 2) the volumetric atrophy analysis. An algorithmic framework was developed to combine both of them and to select the information to be included in textual explanations for the CNN’s predictions (Figure 1).ResultOur framework results in a modular pipeline where each component – attribution method, segmentation tool, ontology model, could be altered due to advancements or changes in the requirements. Our framework implements a plug‐and‐play functionality for interpreting trained CNN models. It assists an easy understanding of CNN’s prediction by generating textual and visual explanations for its findings, with variable degree of detail, focusing on specific ROIs or lobes.ConclusionOur study provides a proof‐of‐concept of utilizing a‐priori neuroanatomical information for explaining CNNs. This prototypical implementation for generating explainable AI is going to be tested by clinicians to evaluate its utility for assisting the diagnosis of dementia. We also plan to extend our framework to other neurodegenerative diseases.

Retrieval of Boundary Layer Precipitable Water from GOES ABI Using Machine Learning Techniques

Abstract Low-level moisture is an important ingredient for forecasting severe storms, especially over the Great Plains where severe storms often develop along the dryline. Although ground-based observation systems such as lidar or radiosondes on weather balloons provide accurate information on low-level moisture, data are provided at limited locations, and the low temporal resolution of radiosondes makes it difficult to track a rapidly developing dryline. Geostationary satellites provide high spatial and temporal observation, but the channels of current geostationary satellites are mostly sensitive to water vapor at mid- to upper levels. However, the split window difference (SWD) between the “clean” window channel (10.3 μm) and the “dirty” window channel (12.3 μm) is commonly used for estimating low-level water vapor. However, this estimation is complicated by surface temperature contributions, dependence on the lapse rate, and nonlinear relationships between SWD and moisture. This study applies machine learning techniques to infer boundary layer precipitable water (BLPW) from Geostationary Operational Environmental Satellite (GOES) Advanced Baseline Imager (ABI) data. Since there are few observations that cover wide regions for training convolutional neural networks, especially for the atmosphere above the surface, High-Resolution Rapid Refresh (HRRR) model outputs are used as the truth for training. Statistical results show that using ABI channel 13, 14, and 15 brightness temperatures, skin temperature, solar zenith angle, and GOES total precipitable water product as additional inputs shows the lowest mean-square error (MSE). Case study results show that the model developed in this study is good at capturing strong moisture gradients and depicting a dryline. Significance Statement Knowledge of low-level moisture is important for monitoring drylines and moisture convergence preceding convective initiation. Traditional observations lack the spatial resolution, spatial coverage, or temporal update frequency needed by forecasters. The latest generation of Geostationary Operational Environmental Satellite (GOES), the GOES-R series, provides high spatial resolution and rapid temporal updates. However, the estimation of low-level moisture using the Advanced Baseline Imager (ABI) is complicated by dependencies on other variables and nonlinear relationships. Machine learning is shown to provide good estimates of low-level moisture from ABI observations. This application of machine learning to GOES-R can improve the ability to monitor low-level moisture and forecast convective initiation.

Image Quality Assessment and Reliability Analysis of Artificial Intelligence-Based Tumor Classification of Stimulated Raman Histology of Tumor Biobank Samples.

Stimulated Raman histology (SRH) is a label-free optical imaging method for rapid intraoperative analysis of fresh tissue samples. Analysis of SRH images using Convolutional Neural Networks (CNN) has shown promising results for predicting the main histopathological classes of neurooncological tumors. Due to the relatively low number of rare tumor representations in CNN training datasets, a valid prediction of rarer entities remains limited. To develop new reliable analysis tools, larger datasets and greater tumor variety are crucial. One way to accomplish this is through research biobanks storing frozen tumor tissue samples. However, there is currently no data available regarding the pertinency of previously frozen tissue samples for SRH analysis. The aim of this study was to assess image quality and perform a comparative reliability analysis of artificial intelligence-based tumor classification using SRH in fresh and frozen tissue samples. In a monocentric prospective study, tissue samples from 25 patients undergoing brain tumor resection were obtained. SRH was acquired in fresh and defrosted samples of the same specimen after varying storage durations at -80 °C. Image quality was rated by an experienced neuropathologist, and prediction of histopathological diagnosis was performed using two established CNNs. The image quality of SRH in fresh and defrosted tissue samples was high, with a mean image quality score of 1.96 (range 1-5) for both groups. CNN analysis showed high internal consistency for histo-(Cα 0.95) and molecular (Cα 0.83) pathological tumor classification. The results were confirmed using a dataset with samples from the local tumor biobank (Cα 0.91 and 0.53). Our results showed that SRH appears comparably reliable in fresh and frozen tissue samples, enabling the integration of tumor biobank specimens to potentially improve the diagnostic range and reliability of CNN prediction tools.

Predicting Cloud-To-Ground Lightning in the Western United States From the Large-Scale Environment Using Explainable Neural Networks.

Lightning is a major source of wildfire ignition in the western United States (WUS). We build and train convolutional neural networks (CNNs) to predict the occurrence of cloud-to-ground (CG) lightning across the WUS during June-September from the spatial patterns of seven large-scale meteorological variables from reanalysis (1995-2022). Individually trained CNN models at each 1°×1° grid cell (n=285 CNNs) show high skill at predicting CG lightning days across the WUS (median AUC=0.8) and perform best in parts of the interior Southwest where summertime CG lightning is most common. Further, interannual correlation between observed and predicted CG lightning days is high (median r=0.87), demonstrating that locally trained CNNs realistically capture year-to-year variation in CG lightning activity across the WUS. We then use layer-wise relevance propagation (LRP) to investigate the relevance of predictor variables to successful CG lightning prediction in each grid cell. Using maximum LRP values, our results show that two thermodynamic variables-ratio of surface moist static energy to free-tropospheric saturation moist static energy, and the 700-500hPa lapse rate-are the most relevant CG lightning predictors for 93%-96% of CNNs depending on the LRP variant used. As lightning is not directly simulated by global climate models, these CNNs could be used to parameterize CG lightning in climate models to assess changes in future CG lightning occurrence with projected climate change. Understanding changes in CG lightning risk and consequently lightning-caused wildfire risk across the WUS could inform fire management, planning, and disaster preparedness.

Incorporating label uncertainty during the training of convolutional neural networks improves performance for the discrimination between certain and inconclusive cases in dopamine transporter SPECT.

Deep convolutional neural networks (CNN) hold promise for assisting the interpretation of dopamine transporter (DAT)-SPECT. For improved communication of uncertainty to the user it is crucial to reliably discriminate certain from inconclusive cases that might be misclassified by strict application of a predefined decision threshold on the CNN output. This study tested two methods to incorporate existing label uncertainty during the training to improve the utility of the CNN sigmoid output for this task. Three datasets were used retrospectively: a "development" dataset (n = 1740) for CNN training, validation and testing, two independent out-of-distribution datasets (n = 640, 645) for testing only. In the development dataset, binary classification based on visual inspection was performed carefully by three well-trained readers. A ResNet-18 architecture was trained for binary classification of DAT-SPECT using either a randomly selected vote ("random vote training", RVT), the proportion of "reduced" votes ( "average vote training", AVT) or the majority vote (MVT) across the three readers as reference standard. Balanced accuracy was computed separately for "inconclusive" sigmoid outputs (within a predefined interval around the 0.5 decision threshold) and for "certain" (non-inconclusive) sigmoid outputs. The proportion of "inconclusive" test cases that had to be accepted to achieve a given balanced accuracy in the "certain" test case was lower with RVT and AVT than with MVT in all datasets (e.g., 1.9% and 1.2% versus 2.8% for 98% balanced accuracy in "certain" test cases from the development dataset). In addition, RVT and AVT resulted in slightly higher balanced accuracy in all test cases independent of their certainty (97.3% and 97.5% versus 97.0% in the development dataset). Making between-readers-discrepancy known to CNN during the training improves the utility of their sigmoid output to discriminate certain from inconclusive cases that might be misclassified by the CNN when the predefined decision threshold is strictly applied. This does not compromise on overall accuracy.

Comparison of Manual vs Artificial Intelligence-Based Muscle MRI Segmentation for Evaluating Disease Progression in Patients With CMT1A.

Intramuscular fat fraction (FF), assessed with quantitative MRI (qMRI), has emerged as one of the few responsive outcome measures in CMT1A patients. The main limitation for its use in future therapeutic trials is the time required for the manual segmentation of individual muscles. This study aimed to evaluate the accuracy and responsiveness of a fully automatic artificial intelligence (AI)-based segmentation pipeline to assess disease progression in a cohort of CMT1A patients over 1 year. Twenty CMT1A patients were included in this observational, prospective, longitudinal study. FF was measured twice a year apart using qMRI in the lower limbs. Individual muscle segmentation was performed fully automatically using a trained convolutional neural network with or without human quality check (QC). The corresponding results were compared with those obtained by fully manual (FM) segmentation using the Dice similarity coefficient (DSC). FF progression and its standardized response mean (SRM) were also computed in individual muscles over the single central slice and a 3D volume to define the most sensitive region of interest. AI-based segmentation showed excellent DSC values (>0.90). Significant global FF progression was observed at thigh (+0.71% ± 1.28%; p = 0.016) and leg (+1.73% ± 2.88%, p = 0.007) levels, similarly to that calculated using the FM technique (p = 0.363 and p = 0.634). FF progression of each individual muscle was comparable when computed from either the central slice or the 3D volume. The best SRM value (0.70) was obtained for the FF progression computed using the AI-based technique with human QC in the 3D volume at the leg level. The time required for fully automatic segmentation using AI with a QC was 10 hours for the entire data set compared with 90 hours for the FM. qMRI combined with AI-based segmentation can be considered as a process ready for assessing longitudinal FF changes in CMT1A patients. Given the slow FF progression at a thigh level and the large heterogeneity between muscles and individuals, FF should be quantified from a 3D volume at the leg level for longitudinal analyses. A QC performed after the AI-based segmentation is still advised given the increased SRM value.