Are Explanations Helpful? A Comparative Analysis of Explainability Methods in Skin Lesion Classifiers

Deep Learning has shown outstanding results in computer vision tasks; healthcare is no exception. However, there is no straightforward way to expose the decision-making process of DL models. Good accuracy is not enough for skin cancer predictions. Understanding the model's behavior is crucial for clinical application and reliable outcomes. In this work, we identify desiderata for explanations in skin-lesion models. We analyzed seven methods, four based on pixel-attribution (Grad-CAM, Score-CAM, LIME, SHAP) and three on high-level concepts (ACE, ICE, CME), for a deep neural network trained on the International Skin Imaging Collaboration Archive. Our findings indicate that while these techniques reveal biases, there is room for improving the comprehensiveness of explanations to achieve transparency in skin-lesion models.

BiasPruner: Debiased Continual Learning for Medical Image Classification

Continual Learning (CL) is crucial for enabling networks to dynamically adapt as they learn new tasks sequentially, accommodating new data and classes without catastrophic forgetting. Diverging from conventional perspectives on CL, our paper introduces a new perspective wherein forgetting could actually benefit the sequential learning paradigm. Specifically, we present BiasPruner, a CL framework that intentionally forgets spurious correlations in the training data that could lead to shortcut learning. Utilizing a new bias score that measures the contribution of each unit in the network to learning spurious features, BiasPruner prunes those units with the highest bias scores to form a debiased subnetwork preserved for a given task. As BiasPruner learns a new task, it constructs a new debiased subnetwork, potentially incorporating units from previous subnetworks, which improves adaptation and performance on the new task. During inference, BiasPruner employs a simple task-agnostic approach to select the best debiased subnetwork for predictions. We conduct experiments on three medical datasets for skin lesion classification and chest X-Ray classification and demonstrate that BiasPruner consistently outperforms SOTA CL methods in terms of classification performance and fairness. Our code is available here.

Back to the Basics on Predicting Transfer Performance

In the evolving landscape of deep learning, selecting the best pre-trained models from a growing number of choices is a challenge. Transferability scorers propose alleviating this scenario, but their recent proliferation, ironically, poses the challenge of their own assessment. In this work, we propose both robust benchmark guidelines for transferability scorers, and a well-founded technique to combine multiple scorers, which we show consistently improves their results. We extensively evaluate 13 scorers from literature across 11 datasets, comprising generalist, fine-grained, and medical imaging datasets. We show that few scorers match the predictive performance of the simple raw metric of models on ImageNet, and that all predictors suffer on medical datasets. Our results highlight the potential of combining different information sources for reliably predicting transferability across varied domains.

Key Patches Are All You Need: A Multiple Instance Learning Framework For Robust Medical Diagnosis

Deep learning models have revolutionized the field of medical image analysis, due to their outstanding performances. However, they are sensitive to spurious correlations, often taking advantage of dataset bias to improve results for in-domain data, but jeopardizing their generalization capabilities. In this paper, we propose to limit the amount of information these models use to reach the final classification, by using a multiple instance learning (MIL) framework. MIL forces the model to use only a (small) subset of patches in the image, identifying discriminative regions. This mimics the clinical procedures, where medical decisions are based on localized findings. We evaluate our framework on two medical applications: skin cancer diagnosis using dermoscopy and breast cancer diagnosis using mammography. Our results show that using only a subset of the patches does not compromise diagnostic performance for in-domain data, compared to the baseline approaches. However, our approach is more robust to shifts in patient demographics, while also providing more detailed explanations about which regions contributed to the decision. Code is available at: https://github.com/diogojpa99/MedicalMultiple-Instance-Learning.

MMIST-ccRCC: A Real World Medical Dataset for the Development of Multi-Modal Systems

The acquisition of different data modalities can enhance our knowledge and understanding of various diseases, paving the way for a more personalized healthcare. Thus, medicine is progressively moving towards the generation of massive amounts of multi-modal data (\emph{e.g,} molecular, radiology, and histopathology). While this may seem like an ideal environment to capitalize data-centric machine learning approaches, most methods still focus on exploring a single or a pair of modalities due to a variety of reasons: i) lack of ready to use curated datasets; ii) difficulty in identifying the best multi-modal fusion strategy; and iii) missing modalities across patients. In this paper we introduce a real world multi-modal dataset called MMIST-CCRCC that comprises 2 radiology modalities (CT and MRI), histopathology, genomics, and clinical data from 618 patients with clear cell renal cell carcinoma (ccRCC). We provide single and multi-modal (early and late fusion) benchmarks in the task of 12-month survival prediction in the challenging scenario of one or more missing modalities for each patient, with missing rates that range from 26$\%$ for genomics data to more than 90$\%$ for MRI. We show that even with such severe missing rates the fusion of modalities leads to improvements in the survival forecasting. Additionally, incorporating a strategy to generate the latent representations of the missing modalities given the available ones further improves the performance, highlighting a potential complementarity across modalities. Our dataset and code are available here: https://multi-modal-ist.github.io/datasets/ccRCC

XAI for Skin Cancer Detection with Prototypes and Non-Expert Supervision

Skin cancer detection through dermoscopy image analysis is a critical task. However, existing models used for this purpose often lack interpretability and reliability, raising the concern of physicians due to their black-box nature. In this paper, we propose a novel approach for the diagnosis of melanoma using an interpretable prototypical-part model. We introduce a guided supervision based on non-expert feedback through the incorporation of: 1) binary masks, obtained automatically using a segmentation network; and 2) user-refined prototypes. These two distinct information pathways aim to ensure that the learned prototypes correspond to relevant areas within the skin lesion, excluding confounding factors beyond its boundaries. Experimental results demonstrate that, even without expert supervision, our approach achieves superior performance and generalization compared to non-interpretable models.

The Performance of Transferability Metrics does not Translate to Medical Tasks

Transfer learning boosts the performance of medical image analysis by enabling deep learning (DL) on small datasets through the knowledge acquired from large ones. As the number of DL architectures explodes, exhaustively attempting all candidates becomes unfeasible, motivating cheaper alternatives for choosing them. Transferability scoring methods emerge as an enticing solution, allowing to efficiently calculate a score that correlates with the architecture accuracy on any target dataset. However, since transferability scores have not been evaluated on medical datasets, their use in this context remains uncertain, preventing them from benefiting practitioners. We fill that gap in this work, thoroughly evaluating seven transferability scores in three medical applications, including out-of-distribution scenarios. Despite promising results in general-purpose datasets, our results show that no transferability score can reliably and consistently estimate target performance in medical contexts, inviting further work in that direction.

Test-Time Selection for Robust Skin Lesion Analysis

Skin lesion analysis models are biased by artifacts placed during image acquisition, which influence model predictions despite carrying no clinical information. Solutions that address this problem by regularizing models to prevent learning those spurious features achieve only partial success, and existing test-time debiasing techniques are inappropriate for skin lesion analysis due to either making unrealistic assumptions on the distribution of test data or requiring laborious annotation from medical practitioners. We propose TTS (Test-Time Selection), a human-in-the-loop method that leverages positive (e.g., lesion area) and negative (e.g., artifacts) keypoints in test samples. TTS effectively steers models away from exploiting spurious artifact-related correlations without retraining, and with less annotation requirements. Our solution is robust to a varying availability of annotations, and different levels of bias. We showcase on the ISIC2019 dataset (for which we release a subset of annotated images) how our model could be deployed in the real-world for mitigating bias.

Even Small Correlation and Diversity Shifts Pose Dataset-Bias Issues

Distribution shifts are common in real-world datasets and can affect the performance and reliability of deep learning models. In this paper, we study two types of distribution shifts: diversity shifts, which occur when test samples exhibit patterns unseen during training, and correlation shifts, which occur when test data present a different correlation between seen invariant and spurious features. We propose an integrated protocol to analyze both types of shifts using datasets where they co-exist in a controllable manner. Finally, we apply our approach to a real-world classification problem of skin cancer analysis, using out-of-distribution datasets and specialized bias annotations. Our protocol reveals three findings: 1) Models learn and propagate correlation shifts even with low-bias training; this poses a risk of accumulating and combining unaccountable weak biases; 2) Models learn robust features in high- and low-bias scenarios but use spurious ones if test samples have them; this suggests that spurious correlations do not impair the learning of robust features; 3) Diversity shift can reduce the reliance on spurious correlations; this is counter intuitive since we expect biased models to depend more on biases when invariant features are missing. Our work has implications for distribution shift research and practice, providing new insights into how models learn and rely on spurious correlations under different types of shifts.

Artifact-Based Domain Generalization of Skin Lesion Models

Deep Learning failure cases are abundant, particularly in the medical area. Recent studies in out-of-distribution generalization have advanced considerably on well-controlled synthetic datasets, but they do not represent medical imaging contexts. We propose a pipeline that relies on artifacts annotation to enable generalization evaluation and debiasing for the challenging skin lesion analysis context. First, we partition the data into levels of increasingly higher biased training and test sets for better generalization assessment. Then, we create environments based on skin lesion artifacts to enable domain generalization methods. Finally, after robust training, we perform a test-time debiasing procedure, reducing spurious features in inference images. Our experiments show our pipeline improves performance metrics in biased cases, and avoids artifacts when using explanation methods. Still, when evaluating such models in out-of-distribution data, they did not prefer clinically-meaningful features. Instead, performance only improved in test sets that present similar artifacts from training, suggesting models learned to ignore the known set of artifacts. Our results raise a concern that debiasing models towards a single aspect may not be enough for fair skin lesion analysis.