Towards Trainable Saliency Maps in Medical Imaging

Aggarwal, Mehak; Arun, Nishanth; Gupta, Sharut; Vaswani, Ashwin; Chen, Bryan; Li, Matthew; Chang, Ken; Patel, Jay; Katherine, Hoebel,; Gidwani, Mishka; Kalpathy-Cramer, Jayashree; Singh, Praveer

doi:10.48550/arxiv.2011.07482

Cited by 2 publications

(2 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…An important example is the domain of biomedical imaging. Here, medical practitioners are often interested in what the most important regions of a radiographic image are for a specific prediction (Arun et al, 2020a,b;Aggarwal et al, 2020), in order to build trust in the model and identify when a model might be making a mistake. Application of the general guidelines given above can help developers choose the right attribution method in this case.…”

Section: Discussionmentioning

confidence: 99%

Evaluating Feature Attribution Methods in the Image Domain

Gevaert¹,

Rousseau²,

Becker³

et al. 2022

Preprint

View full text Add to dashboard Cite

Feature attribution maps are a popular approach to highlight the most important pixels in an image for a given prediction of a model. Despite a recent growth in popularity and available methods, little attention is given to the objective evaluation of such attribution maps. Building on previous work in this domain, we investigate existing metrics and propose new variants of metrics for the evaluation of attribution maps. We confirm a recent finding that different attribution metrics seem to measure different underlying concepts of attribution maps, and extend this finding to a larger selection of attribution metrics. We also find that metric results on one dataset do not necessarily generalize to other datasets, and methods with desirable theoretical properties such as DeepSHAP do not necessarily outperform computationally cheaper alternatives. Based on these findings, we propose a general benchmarking approach to identify the ideal feature attribution method for a given use case. Implementations of attribution metrics and our experiments 1 are available online.

show abstract

Section: Discussionmentioning

confidence: 99%

Evaluating Feature Attribution Methods in the Image Domain

Gevaert¹,

Rousseau²,

Becker³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…However, this data collection method is impractical, since the whole process may take too long a period, be expensive, and might get imbalanced distributions among different diseases. Differently, radiologists and computer scientists have recently managed to use saliency methods [44][45][46][47] , to generate heat maps that highlight the regions significantly contributing to the diagnosis, and used the heat-maps as labeling of the CheXpert dataset 48 , a large public available dataset. Saliency methods offer post-hoc interpretability for the pathological regions in the radiographs but are not exposed to pixel-level localization 49 .…”

mentioning

confidence: 99%

Prompt-driven Healthy/Diseased Image Pairs Enabling Pixel-level Chest X-ray Pathology Localization

Suo,

Dong,

Cheng

et al. 2024

Preprint

View full text Add to dashboard Cite

Medical artificial intelligence (AI) offers great potential for automatic pathology interpretation, but the performance is far behind providing a practical tool in clinical settings, which demands both pixel-level accuracy and high interpretability for diagnosis. The main challenges lie in that the construction of such AI models relies on substantial training data with fine-grained labeling that is impractical in real applications. To circumvent this barrier, we propose a prompt-driven constrained generative model to produce anatomically aligned healthy/diseased image pairs and then learn a superb pathology localization model in a supervised manner. The new paradigm effectively addresses the lack of high-fidelity chest X-ray images with pathology labeling at fine scales. Benefiting from the emerging text-driven generative foundation model and the newly incorporated constraint, our model presents promising localization accuracy of the subtle pathologies, high interpretability for clinical decisions, and good transfer ability to many unseen pathological categories (e.g., new prompts and mixed pathologies). These advantageous features establish our model as a promising solution to assist chest X-ray analysis. Besides, the proposed approach is also inspiring for other tasks lacking massive training data and time-consuming manual labeling.

show abstract

Towards Trainable Saliency Maps in Medical Imaging

Cited by 2 publications

References 0 publications

Evaluating Feature Attribution Methods in the Image Domain

Evaluating Feature Attribution Methods in the Image Domain

Prompt-driven Healthy/Diseased Image Pairs Enabling Pixel-level Chest X-ray Pathology Localization

Contact Info

Product

Resources

About