Abstract PO-024: Boosting up knee bone tumor detection from radiology and magnetic resonance imaging by using deep learning techniques

Abstract

Abstract Bone tumors can occur at all ages and also appear around skeletal positions of the human knee. Since the types and expression patterns are extremely diverse, some cases are non-specific symptoms or the onset of symptoms is delayed, and the diagnosis is delayed by low suspicion of physicians. However, timely diagnosis is very important for treatment to reduce extreme pain for patients. Therefore, detecting tumors from radiography and MRI images based on deep learning has recently emerged with the advantage of remarkably shortening time and a completely automatic diagnosis. In this study, we developed a radiology and MRI dataset for knee bone tumor and contributed a robust deep learning-based method to detect tumors. Through an association with Chonnam National University Hospital, we built the radiology knee bone tumor dataset with 1695 high-resolution images including normal, benign and malignant. It is the first dataset for knee bone tumor detection with five categories for classification, and tumor masks, bone regions for segmentation. We proposed multi-level distance features to exploit the tumors based on neighboring regions dealing with challenges from small size, appearance variety, and uncommon. It led to interpreting high-risk regions occurring tumors. Multi-task learning with classification and segmentation helped to learn the mutual benefit from the pixel-wise segmentation and image classification. Our model is fined-tuned in image patches to boost up the accuracy in malignant classification. In difficult cases, MRI has been proved to be a preeminent option to be capable of presenting bone marrow and soft tissue. Therefore, we built MRI knee bone tumor dataset with 104 study cases over 900 series. We used Slicer3D for making the ground-truth in T1-, T2-, and PD- weighted MRI. We generated tumor and bone segmentation masks and classified the study case into three groups: normal, benign, and malignant. We made the ablation study experiments to our radiology dataset. For multi-task learning, we increased Mean IoU from 69.5% to 77.3%. Multi-level distance features helped to increase to 78.8%. For classification, multi-task learning boosted the accuracy from 93% to 95.3%, and multi-level distance features helped the increase of 3%. Besides, the patching training improved malignant classification from 66.7% to 81.5%. The results demonstrated that our deep learning-based method made an important contribution to help doctors in the clinical diagnosis process using a radiology image. Furthermore, our method produced the high-risk neighbor tumor regions by multi-level distance features instead of extracting exactly tumor regions. To improve more accuracy in automatic detection, we hope that MRI images with high contrast resolution can describe clearly soft tissue components and help to recognize the attributes of the tumors clearly. Nonetheless, the dataset of knee bone tumor on MRI is raw and not standard with high complexity. Preprocessing these MRI images is also our initial success to the steppingstone of further research. Citation Format: Nhu-Tai Do, Sung-Taek Jung, Soo-Hyung Kim. Boosting up knee bone tumor detection from radiology and magnetic resonance imaging by using deep learning techniques [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-024.