Deep Omni-supervised Learning for Rib Fracture Detection from Chest Radiology Images

TL;DR

Rib fracture detection in chest radiology is hampered by the need for fine-grained annotations. This paper introduces ORF-Netv2, an omni-supervised detector with a multi-branch head and co-training-guided dynamic label assignment that learns from box, mask, dot, and unlabeled data without relying on pseudo bounding boxes. The method leverages inter-guided maps to weight samples and integrates per-data-type losses in a shared localization framework, with extensive experiments on three datasets showing consistent gains over box-only baselines and competitive label-efficient approaches. The framework also explores budget-aware labeling policies, demonstrating practical guidance for annotation strategies and broad applicability to other medical detection tasks.

Abstract

Deep learning (DL)-based rib fracture detection has shown promise of playing an important role in preventing mortality and improving patient outcome. Normally, developing DL-based object detection models requires a huge amount of bounding box annotation. However, annotating medical data is time-consuming and expertise-demanding, making obtaining a large amount of fine-grained annotations extremely infeasible. This poses a pressing need {for} developing label-efficient detection models to alleviate radiologists' labeling burden. To tackle this challenge, the literature on object detection has witnessed an increase of weakly-supervised and semi-supervised approaches, yet still lacks a unified framework that leverages various forms of fully-labeled, weakly-labeled, and unlabeled data. In this paper, we present a novel omni-supervised object detection network, ORF-Netv2, to leverage as much available supervision as possible. Specifically, a multi-branch omni-supervised detection head is introduced with each branch trained with a specific type of supervision. A co-training-based dynamic label assignment strategy is then proposed to enable flexible and robust learning from the weakly-labeled and unlabeled data. Extensive evaluation was conducted for the proposed framework with three rib fracture datasets on both chest CT and X-ray. By leveraging all forms of supervision, ORF-Netv2 achieves mAPs of 34.7, 44.7, and 19.4 on the three datasets, respectively, surpassing the baseline detector which uses only box annotations by mAP gains of 3.8, 4.8, and 5.0, respectively. Furthermore, ORF-Netv2 consistently outperforms other competitive label-efficient methods over various scenarios, showing a promising framework for label-efficient fracture detection. The code is available at: https://github.com/zhizhongchai/ORF-Net.

Figures (5)

• Figure 1: Medical data can have fine-grained annotations, such as (a) boxes and (b) masks, coarse-grained annotations (or weak annotations), such as (c) dots, and the data are more often (d) unlabeled. Rib fractures are highlighted by zooming in.
• Figure 2: Schematic view of our proposed framework. The network consists of a Feature Pyramid Network (FPN lin2017feature) as the backbone, and an omni-supervised detection head to predict the classification score and localization information. For each form of annotated data, there is a corresponding classification branch that is trained using a dynamic label assignment strategy.
• Figure 3: Visualization of the predicted maps from different classification branches at different iteration numbers.
• Figure 4: Qualitative comparisons of the FCOS tian2020fcos, UT liu2021unbiased, ORF-Net chai2022orf, and our proposed method on RibFrac and XRF. Ground truth, true positives, and false positives are annotated in red boxes, green boxes, and blue boxes, respectively.
• Figure 5: Visualization of the prediction probability maps with different label assignment strategies.