A Systematic framework for optimizing kidney stone detection in low data medical setting

n medical imaging, kidney stone detection is a
crucial task where precise and effective identification may have a
big impact on patient treatment. The lack of annotated medical
datasets and their expensive cost frequently restrict the use of
traditional supervised algorithms. In order to increase perfor-
mance and resilience in a low-data setting, this research explores
the combination of revolutionary YOLO detectors with self-
supervised learning (SSL) methods. Using a single-class kidney
stone dataset, we assess YOLOv10, YOLOv11, and YOLOv12 as
baseline models. With a test mAP@0.5 of 0.7170, we find that
YOLOv11 is the better baseline. Next, using the same dataset, we
pretrain the YOLOv11 backbone using three SSL frameworks:
SimCLR, BYOL and DINO. In particular, the test mAP@0.5
is improved to 0.7734 and mAP@0.5:0.95 to 0.3571 by DINO
pretraining, indicating improved localization accuracy. This study
also presents an end-to-end real-time vision transformer called
RF-DETR Nano. RF-DETR Nano achieves a validation mAP@0.5
of 0.7561 and a peak test mAP@0.5 of 0.8132 by avoiding
conventional Non-Maximum Suppression (NMS) and utilizing
receptive-field attention. This work demonstrates that SSL and
transformer-based detection heads offers a strong method for
enhancing object recognition in medical imaging, providing a
reliable solution in situations when there is a shortage of labeled
data.