Content-Based Medical Image Retrieval for Medical Radiology Images

In this study, a CBIR system for medical radiology images of various anatomical regions and imaging modalities was designed and implemented. Three descriptor-extraction retrieval methods were evaluated: BoVW, HOG, and CNN. Optimal configuration for each method was selected empirically. We used a dataset consisting of 38, 000 2D images, originating from CHC Rijeka, Croatia. Retrieval with ImageNet-pretrained CNNs outperformed both BoVW and HOG retrieval. Re-ranking retrieved images using a different method improved results in almost all cases. Out of all methods compared, CNN retrieval (ShuffleNet combined with average pooling) with HOG re-ranking performed best. Our experiments can be recreated using the code available at https://github.com/dbarac/cbmir .