Osteomyelitis of the jaw (OMJ) is a serious bacterial infection that affects the jaw bones, which can be caused by a variety of bacteria. Osteonecrosis of the jaw (ONJ) is bone necrosis following a bacterial infection. Symptoms may include pain, swelling, redness, fever, and bone exposure. Treatment typically involves antibiotics to kill the bacteria and surgery to remove all necrotic bone, and recent researches suggested that the early surgical resection improves the prognosis. However, it is difficult to determine the necrotic bone in pre-operative CT images. A non-invasive ONJ region recognition system is required for the precise pre-operative surgical planning, and also it has possibility of early diagnosis. This study proposes a ONJ region detection method using contrastive learning (CL) and support vector machine. The proposed method trains the CNN for feature extraction according to the CL framework, and the extracted feature is fed into support vector machine to perform two-class (normal or ONJ) classification on each subdivided image patch. In the experiment, we compared patch- and label-based CL with two types of patch extraction criteria. The proposed method was validated on 3D CT images of nine subjects, and the patch-based CL extracted ONJ region in the highest F1 of 0.734. Experimental results also suggested that the smaller stride should improve detection accuracy, and it has applicability to pixel-level ONJ detection.
View full abstract