# Weakly supervised multiple-instance active learning for tooth-marked tongue recognition

## TL;DR

This paper introduces a new method for identifying tooth marks on tongues in traditional Chinese medicine using less manual labeling, improving efficiency and accuracy.

## Contribution

A weakly supervised multiple-instance active learning model is proposed to reduce annotation workload for tooth-marked tongue recognition.

## Key findings

- The method achieves 93.88% accuracy in tooth-marked tongue recognition.
- It performs comparably to methods requiring more detailed annotations with significantly less labeling effort.
- The proposed approach improves training efficiency by using pseudo-labeled unlabeled data.

## Abstract

Recognizing a tooth-marked tongue has important clinical diagnostic value in traditional Chinese medicine. Current deep learning methods for tooth mark detection require extensive manual labeling and tongue segmentation, which is labor-intensive. Therefore, we propose a weakly supervised multipleinstance active learning model for tooth-marked tongue recognition, aiming to eliminate preprocessing segmentation and reduce the annotation workload while maintaining diagnostic accuracy.

We propose a one-stage method tongenerate tooth mark instances that eliminates the need for pre-segmentation of the tongue. To make full use of unlabeled data, we introduce a semisupervised learning paradigm to pseudo-label unlabeled tongue images with high model confidence in active learning and incorporate them into the training set to improve the training efficiency of the active learning model. In addition, we propose an instance-level hybrid query method considering the diversity of tooth marks.

Experimental results on clinical tongue images verify the effectiveness of the proposed method, which achieves an accuracy of 93.88% for tooth-marked tongue recognition, outperforming the recently introduced weakly supervised approaches.

The proposed method is effective with only a small amount of image-level annotation, and its performance is comparable to that of image-level annotation, instance-level annotation and pixel-level annotation, which require a large number of tooth markers. Our method significantly reduces the annotation cost of the binary classification task of traditional Chinese medicine tooth mark recognition.