DAL: Dynamic Angular Loss for Imbalanced Medical Image Classification

Salman Mohammad, Furkan Kasım, Manuel Günther

Keywords: Image Classification, Class Imbalance, Angular Loss

Abstract:

Class imbalance remains a major obstacle in medical image classification, where rare but clinically important classes are often overshadowed by majority categories.
Despite current strategies for reducing bias, including angular-margin losses such as ArcFace, offer strong discriminative features, their behavior under small amounts of severely imbalanced classes is insufficiently understood.
Although previous approaches estimate intrinsic parameters of these losses, their assumptions are based on large amounts of classes, which do not translate into medical cases.
We introduce a dynamic angular loss (DAL) to generalize these parameters using analytically derived negative-angle statistics, and combine it with a batch-adaptive angular margin and dynamically-weighted cross-entropy.
Across three highly imbalanced medical image classification benchmarks, our method consistently achieves superior balanced accuracy scores, and the lowest coefficient of variation of per-class recall.
Although this comes with a small performance reduction on majority classes, gains on minority categories are substantial, resulting in a reliable and equitable overall classifier.

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Fairness and Bias

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 100