Automatic detecting multiple bone metastases in breast cancer using deep learning based on low-resolution bone scan images

Whole-body bone scan (WBS) is usually used as the effective diagnostic method for early-stage and comprehensive bone metastases of breast cancer. WBS images with breast cancer bone metastasis have the characteristics of low resolution, small foreground, and multiple lesions, hindering the widespread application of deep learning-based models. Automatically detecting a large number of densely small lesions on low-resolution WBS images remains a challenge. We aim to develop a unified framework for detecting multiple densely bone metastases based on low-resolution WBS images. We propose a novel unified detection framework to detect multiple bone metastases based on WBS images. Considering the difficulties of feature extraction caused by low resolution and multiple lesions, we innovatively propose the plug-and-play position auxiliary extraction module and feature fusion module to enhance the ability of global information extraction. In order to accurately detect small metastases in WBS, we designed the self-attention transformer-based target detection head. This retrospective study included 512 patients with breast cancer bone metastases from Peking Union Medical College Hospital. The data type is whole-body bone scan image. For our study, the ratio of training set, validation set and test set is about 6:2:2. The benchmarks are four representative baselines, SSD, YOLOR, Faster_RCNN_R and Scaled-YOLOv4. The performance metrics are Average Precision (AP), Precision and Recall. The detection results obtained through the proposed method were assessed using the Bonferroni-adjusted Wilcoxon rank test. The significant level is adjusted according to different multiple comparisons. We conducted extensive experiments and ablation studies on a private dataset of breast cancer WBS and a public dataset of bone scans from West China Hospital to validate the effectiveness and generalization. Experiments were conducted to evaluate the effectiveness of our method. First, compared to different network architectures, our method obtained AP of 55.0 ± 6.4% (95% confidence intervals (CI) 49.9-60.1%, [Formula: see text]), which improved AP by 45.2% for the SSD baseline with AP 9.8 ± 2% (95% CI 8.1-11.4%). For the metric of recall, our method achieved the average of 54.3 ± 4.2% (95% CI 50.9-57.6%, [Formula: see text]), which has improved the recall values by 49.01% for the SSD model with 5.2 ± 12.7% (95% CI 10-21.3%). Second, we conducted ablation studies. On the private dataset, adding the detection head module and position auxiliary extraction module will increase the AP values by 14.03% (from 33.3 ± 2% to 47.6 ± 4.4%) and 19.3% (from 33.3 ± 2% to 52.6 ± 6.1%), respectively. In addition, the generalization of the method was also verified on the public dataset BS-80K from West China Hospital. Extensive experimental results have demonstrated the superiority and effectiveness of our method. To the best of our knowledge, our work is the first attempt for developing automatic detector considering the unique characteristics of low resolution, small foreground and multiple lesions of breast cancer WBS images. Our framework is tailored for whole-body WBS and can be used as a clinical decision support tool for early decision-making for breast cancer bone metastases.