Explainable multiple abnormality classification of chest CT volumes.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Explainable AI (XAI)
• Computer-Aided Diagnosis

Background:

• Accurate interpretation of medical images is crucial for patient care and requires understanding AI model predictions.
• Existing models may exploit confounding variables, leading to incorrect diagnoses.
• Explainable AI is needed to verify model correctness and identify regions used for predictions.

Purpose of the Study:

• To introduce explainable multiple abnormality classification in volumetric medical images.
• To propose a novel AI model, AxialNet, capable of identifying abnormalities and their locations.
• To improve the accuracy of abnormality prediction and organ localization in chest CT scans.

Main Methods:

• Developed AxialNet, a multiple instance learning convolutional neural network for abnormality identification.
• Incorporated HiResCAM, an attention mechanism, for precise sub-slice region identification.
• Introduced a novel mask loss function combined with PARTITION (radiology report and image processing) for organ-specific abnormality localization.

Main Results:

• AxialNet with HiResCAM provides faithful explanations, unlike Grad-CAM.
• The mask loss improved organ localization of multiple abnormalities by 33% on the RAD-ChestCT dataset.
• Achieved state-of-the-art performance in explainable multi-abnormality prediction in chest CT volumes.

Conclusions:

• AxialNet is the first model for explainable multi-abnormality prediction in volumetric medical images.
• The proposed methods significantly enhance the accuracy and reliability of AI in medical diagnostics.
• This work advances the clinical applicability of AI for chest CT volume analysis.