[1M3-CC] Machine Learning Application to Medical Engineering

[1M3-CC-03] Automatic benign and malignant estimation of bone tumors using deep learning

Kaito Furuo ¹, Kento Morita ¹, Tomohito Hagi ², Tomoki Nakamura ², Tetsushi Wakabayashi ¹ (1. Mie University, 2. Mie University Hospital)

The bone tumor causes the bone pain and swelling, and is firstly diagnosed in a local hospital in many cases. This has become a problem in recent years, and also the benign and malignant nature of bone tumors is difficult and requires a great deal of effort even for medical specialists. Therefore, the development of a system to automatically estimate the benign or malignant nature of bone tumors is required. In this study, we propose a method for automatically estimating the benignity or malignancy of bone tumors using deep learning. We fine-tuned VGG16 and ResNet152 trained on ImageNet using image patches extracted from 38 plain X-ray images of 3 patients. Results on patch-level classification showed that VGG16 achieved higher estimation accuracy (f1-score of 0.790) than ResNet152 (f1- score of 0.784). We also performed the tumor-level classification experiment in which 4 benign and 6 malignant tumors were correctly classified to the appropriate class.