Expert-Adaptive Medical Image Segmentation

TL;DR

This paper tackles the problem of expert-specific annotation bias in medical image segmentation (MIS) by proposing an expert-adaptive MIS pipeline that uses multi-expert annotations to train a conditioned instance normalization (CIN) based multi-task DNN with a shared backbone and expert-specific branches. A new expert is accommodated through lightweight fine-tuning of a newly initialized expert head while keeping the shared backbone fixed, optimizing the loss $l_{ft}$ with a small annotated set $N_{ft}$. Experiments on brain MRI data from the QUBIQ brain-growth dataset show that cross-expert models suffer notable performance drops, while the proposed method improves adaptation with limited data, particularly when the new expert provides fewer than 20 annotations; gains persist as the number of training experts increases but may plateau with larger annotation sets. The work highlights the practical potential of rapid expert adaptation in MIS, reducing annotation burden and enabling deployment across diverse clinical annotators, with avenues for further optimization through alternative training strategies and optimization techniques.

Abstract

Medical image segmentation (MIS) plays an instrumental role in medical image analysis, where considerable effort has been devoted to automating the process. Currently, mainstream MIS approaches are based on deep neural networks (DNNs), which are typically trained on a dataset with annotations produced by certain medical experts. In the medical domain, the annotations generated by different experts can be inherently distinct due to complexity of medical images and variations in expertise and post-segmentation missions. Consequently, the DNN model trained on the data annotated by some experts may hardly adapt to a new expert. In this work, we evaluate a customised expert-adaptive method, characterised by multi-expert annotation, multi-task DNN-based model training, and lightweight model fine-tuning, to investigate model's adaptivity to a new expert in the situation where the amount and mobility of training images are limited. Experiments conducted on brain MRI segmentation tasks with limited training data demonstrate its effectiveness and the impact of its key parameters.

Figures (2)

• Figure 1: Illustration of the brain contours results created by seven different experts on the same MRI. It can be observed that the regions of interest (ROIs) marked out by different experts are largely consistent, though the contours themselves may vary greatly in different locations. Such contour variations may not impact disease diagnosis, which depends more on features within the ROIs, but may significantly influence the segmentation model trained by using them as annotations.
• Figure 2: An illustration of the customised expert-adaptive medical image segmentation method, composed of multi-expert annotation, multi-task DNN-based model training, and lightweight model fine-tuning (based on a small number of annotated data from a new expert).