Speaker
Description
Mammography screening produces multiple images per patient, each assessed for cancer risk and breast density using the BI-RADS system. While these tasks are clinically related, density both increases cancer risk and reduces imaging sensitivity, multi-task learning models do not consistently outperform single-task baselines. This raises a natural question: how much information is truly shared between the two tasks at the level of learned neural network representations, and can explicitly separating shared from task-specific structure lead to improved predictions?
We investigate this question using Angle-Based Joint and Individual Variation Explained (AJIVE), a statistical method for decomposing shared and task-specific variation. Separate ResNet-based classifiers are trained for cancer risk and density prediction on the VinDr-Mammo dataset. Their learned feature representations are then decomposed into joint components and task-specific components. Our hypothesis being that cancer prediction may improve when using only the cancer-specific component, i.e., after removing density-related structure from the representation.
Preliminary results do not show a significant advantage of disentangled representations. This suggests either that the shared structure between the tasks is more deeply entangled than expected, or that current decomposition methods do not align well with task-relevant features in deep networks.
To interpret the decomposition in the image domain, we compare standard Grad-CAM saliency maps with a weighted variant informed by AJIVE loadings. This allows us to assess whether joint and individual components correspond to distinct anatomical or pathological regions. Alignment with annotated findings bounding boxes in VinDr-Mammo provides a clinically grounded evaluation.