On dataset transferability in medical image classification

Current transferability estimation methods designed for natural image datasets are often suboptimal in medical image classification. These methods primarily focus on estimating the suitability of pre-trained source model features for a target dataset, which can lead to unrealistic predictions, such as suggesting that the target dataset is the best source for itself. To address this, we propose a novel transferability metric that combines feature quality with gradients to evaluate both the suitability and adaptability of source model features for target tasks. We evaluate our approach in two new scenarios: source dataset transferability for medical image classification and cross-domain transferability. Our results show that our method outperforms existing transferability metrics in both settings. We also provide insight into the factors influencing transfer performance in medical image classification, as well as the dynamics of cross-domain transfer from natural to medical images. Additionally, we provide ground-truth transfer performance benchmarking results to encourage further research into transferability estimation for medical image classification. Our code and experiments are available at https://github.com/DovileDo/transferability-in-medical-imaging.

SeagrassFinder: Deep Learning for Eelgrass Detection and Coverage Estimation in the Wild

Seagrass meadows play a crucial role in marine ecosystems, providing important services such as carbon sequestration, water quality improvement, and habitat provision. Monitoring the distribution and abundance of seagrass is essential for environmental impact assessments and conservation efforts. However, the current manual methods of analyzing underwater video transects to assess seagrass coverage are time-consuming and subjective. This work explores the use of deep learning models to automate the process of seagrass detection and coverage estimation from underwater video data. A dataset of over 8,300 annotated underwater images was created, and several deep learning architectures, including ResNet, InceptionNetV3, DenseNet, and Vision Transformer, were evaluated for the task of binary classification of ``Eelgrass Present'' and ``Eelgrass Absent'' images. The results demonstrate that deep learning models, particularly the Vision Transformer, can achieve high performance in predicting eelgrass presence, with AUROC scores exceeding 0.95 on the final test dataset. The use of transfer learning and the application of the Deep WaveNet underwater image enhancement model further improved the models' capabilities. The proposed methodology allows for the efficient processing of large volumes of video data, enabling the acquisition of much more detailed information on seagrass distributions compared to current manual methods. This information is crucial for environmental impact assessments and monitoring programs, as seagrasses are important indicators of coastal ecosystem health. Overall, this project demonstrates the value that deep learning can bring to the field of marine ecology and environmental monitoring.

Mask of truth: model sensitivity to unexpected regions of medical images

The development of larger models for medical image analysis has led to increased performance. However, it also affected our ability to explain and validate model decisions. Models can use non-relevant parts of images, also called spurious correlations or shortcuts, to obtain high performance on benchmark datasets but fail in real-world scenarios. In this work, we challenge the capacity of convolutional neural networks (CNN) to classify chest X-rays and eye fundus images while masking out clinically relevant parts of the image. We show that all models trained on the PadChest dataset, irrespective of the masking strategy, are able to obtain an Area Under the Curve (AUC) above random. Moreover, the models trained on full images obtain good performance on images without the region of interest (ROI), even superior to the one obtained on images only containing the ROI. We also reveal a possible spurious correlation in the Chaksu dataset while the performances are more aligned with the expectation of an unbiased model. We go beyond the performance analysis with the usage of the explainability method SHAP and the analysis of embeddings. We asked a radiology resident to interpret chest X-rays under different masking to complement our findings with clinical knowledge. Our code is available at https://github.com/TheoSourget/MMC_Masking and https://github.com/TheoSourget/MMC_Masking_EyeFundus

Confidence intervals uncovered: Are we ready for real-world medical imaging AI?

Medical imaging is spearheading the AI transformation of healthcare. Performance reporting is key to determine which methods should be translated into clinical practice. Frequently, broad conclusions are simply derived from mean performance values. In this paper, we argue that this common practice is often a misleading simplification as it ignores performance variability. Our contribution is threefold. (1) Analyzing all MICCAI segmentation papers (n = 221) published in 2023, we first observe that more than 50% of papers do not assess performance variability at all. Moreover, only one (0.5%) paper reported confidence intervals (CIs) for model performance. (2) To address the reporting bottleneck, we show that the unreported standard deviation (SD) in segmentation papers can be approximated by a second-order polynomial function of the mean Dice similarity coefficient (DSC). Based on external validation data from 56 previous MICCAI challenges, we demonstrate that this approximation can accurately reconstruct the CI of a method using information provided in publications. (3) Finally, we reconstructed 95% CIs around the mean DSC of MICCAI 2023 segmentation papers. The median CI width was 0.03 which is three times larger than the median performance gap between the first and second ranked method. For more than 60% of papers, the mean performance of the second-ranked method was within the CI of the first-ranked method. We conclude that current publications typically do not provide sufficient evidence to support which models could potentially be translated into clinical practice.

Dataset Distribution Impacts Model Fairness: Single vs. Multi-Task Learning

The influence of bias in datasets on the fairness of model predictions is a topic of ongoing research in various fields. We evaluate the performance of skin lesion classification using ResNet-based CNNs, focusing on patient sex variations in training data and three different learning strategies. We present a linear programming method for generating datasets with varying patient sex and class labels, taking into account the correlations between these variables. We evaluated the model performance using three different learning strategies: a single-task model, a reinforcing multi-task model, and an adversarial learning scheme. Our observations include: 1) sex-specific training data yields better results, 2) single-task models exhibit sex bias, 3) the reinforcement approach does not remove sex bias, 4) the adversarial model eliminates sex bias in cases involving only female patients, and 5) datasets that include male patients enhance model performance for the male subgroup, even when female patients are the majority. To generalise these findings, in future research, we will examine more demographic attributes, like age, and other possibly confounding factors, such as skin colour and artefacts in the skin lesions. We make all data and models available on GitHub.

Exploring connections of spectral analysis and transfer learning in medical imaging

In this paper, we use spectral analysis to investigate transfer learning and study model sensitivity to frequency shortcuts in medical imaging. By analyzing the power spectrum density of both pre-trained and fine-tuned model gradients, as well as artificially generated frequency shortcuts, we observe notable differences in learning priorities between models pre-trained on natural vs medical images, which generally persist during fine-tuning. We find that when a model's learning priority aligns with the power spectrum density of an artifact, it results in overfitting to that artifact. Based on these observations, we show that source data editing can alter the model's resistance to shortcut learning.

Source Matters: Source Dataset Impact on Model Robustness in Medical Imaging

Transfer learning has become an essential part of medical imaging classification algorithms, often leveraging ImageNet weights. However, the domain shift from natural to medical images has prompted alternatives such as RadImageNet, often demonstrating comparable classification performance. However, it remains unclear whether the performance gains from transfer learning stem from improved generalization or shortcut learning. To address this, we investigate potential confounders -- whether synthetic or sampled from the data -- across two publicly available chest X-ray and CT datasets. We show that ImageNet and RadImageNet achieve comparable classification performance, yet ImageNet is much more prone to overfitting to confounders. We recommend that researchers using ImageNet-pretrained models reexamine their model robustness by conducting similar experiments. Our code and experiments are available at https://github.com/DovileDo/source-matters.

Towards actionability for open medical imaging datasets: lessons from community-contributed platforms for data management and stewardship

Medical imaging datasets are fundamental to artificial intelligence (AI) in healthcare. The accuracy, robustness and fairness of diagnostic algorithms depend on the data (and its quality) on which the models are trained and evaluated. Medical imaging datasets have become increasingly available to the public, and are often hosted on Community-Contributed Platforms (CCP), including private companies like Kaggle or HuggingFace. While open data is important to enhance the redistribution of data's public value, we find that the current CCP governance model fails to uphold the quality needed and recommended practices for sharing, documenting, and evaluating datasets. In this paper we investigate medical imaging datasets on CCPs and how they are documented, shared, and maintained. We first highlight some differences between medical imaging and computer vision, particularly in the potentially harmful downstream effects due to poor adoption of recommended dataset management practices. We then analyze 20 (10 medical and 10 computer vision) popular datasets on CCPs and find vague licenses, lack of persistent identifiers and storage, duplicates and missing metadata, with differences between the platforms. We present "actionability" as a conceptual metric to reveal the data quality gap between characteristics of data on CCPs and the desired characteristics of data for AI in healthcare. Finally, we propose a commons-based stewardship model for documenting, sharing and maintaining datasets on CCPs and end with a discussion of limitations and open questions.

Data usage and citation practices in medical imaging conferences

Medical imaging papers often focus on methodology, but the quality of the algorithms and the validity of the conclusions are highly dependent on the datasets used. As creating datasets requires a lot of effort, researchers often use publicly available datasets, there is however no adopted standard for citing the datasets used in scientific papers, leading to difficulty in tracking dataset usage. In this work, we present two open-source tools we created that could help with the detection of dataset usage, a pipeline \url{https://github.com/TheoSourget/Public_Medical_Datasets_References} using OpenAlex and full-text analysis, and a PDF annotation software \url{https://github.com/TheoSourget/pdf_annotator} used in our study to manually label the presence of datasets. We applied both tools on a study of the usage of 20 publicly available medical datasets in papers from MICCAI and MIDL. We compute the proportion and the evolution between 2013 and 2023 of 3 types of presence in a paper: cited, mentioned in the full text, cited and mentioned. Our findings demonstrate the concentration of the usage of a limited set of datasets. We also highlight different citing practices, making the automation of tracking difficult.

Augmenting Chest X-ray Datasets with Non-Expert Annotations

The advancement of machine learning algorithms in medical image analysis requires the expansion of training datasets. A popular and cost-effective approach is automated annotation extraction from free-text medical reports, primarily due to the high costs associated with expert clinicians annotating chest X-ray images. However, it has been shown that the resulting datasets are susceptible to biases and shortcuts. Another strategy to increase the size of a dataset is crowdsourcing, a widely adopted practice in general computer vision with some success in medical image analysis. In a similar vein to crowdsourcing, we enhance two publicly available chest X-ray datasets by incorporating non-expert annotations. However, instead of using diagnostic labels, we annotate shortcuts in the form of tubes. We collect 3.5k chest drain annotations for CXR14, and 1k annotations for 4 different tube types in PadChest. We train a chest drain detector with the non-expert annotations that generalizes well to expert labels. Moreover, we compare our annotations to those provided by experts and show "moderate" to "almost perfect" agreement. Finally, we present a pathology agreement study to raise awareness about ground truth annotations. We make our annotations and code available.