There is a huge gap between the paper and landing of AI on medical. The fact is that some paper just publish toy model with off-the-shelf tool (all kinds of CNN model and opencv) to run the accuracy and metrics without understanding of the details. These models are hard to extend to real scenarios of images with local complex ROI. Resort to some top conference like CVPR and MICCAI as well as some top journals in medical diagnosis. There is also standard open dataset online with order of 10GB and 100GB. If your model can handle that, then your research make sense.