2021/08/16 by Chetan L. Srinidhi, Srinidhi, Chetan L, Anne L. Martel +1 · 1 citation
Computer Science · Medicine · #AI in cancer detection #Cervical Cancer and HPV Research #Computer Vision and Pattern Recognition (cs.CV) #Digital Imaging for Blood Diseases #FOS: Computer and information sciences #Machine Learning (cs.LG)
paper · pdf · doi:10.48550/arxiv.2108.07183
openalex publication_date 2021/08/16 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Self-supervised learning (SSL) has recently shown tremendous potential to\nlearn generic visual representations useful for many image analysis tasks.\nDespite their notable success, the existing SSL methods fail to generalize to\ndownstream tasks when the number of labeled training instances is small or if\nthe domain shift between the transfer domains is significant. In this paper, we\nattempt to improve self-supervised pretrained representations through the lens\nof curriculum learning by proposing a hardness-aware dynamic curriculum\nlearning (HaDCL) approach. To improve the robustness and generalizability of\nSSL, we dynamically leverage progressive harder examples via easy-to-hard and\nhard-to-very-hard samples during mini-batch downstream fine-tuning. We discover\nthat by progressive stage-wise curriculum learning, the pretrained\nrepresentations are significantly enhanced and adaptable to both in-domain and\nout-of-domain distribution data.\n We performed extensive validation on three histology benchmark datasets on\nboth patch-wise and slide-level classification problems. Our curriculum based\nfine-tuning yields a significant improvement over standard fine-tuning, with a\nminimum improvement in area-under-the-curve (AUC) score of 1.7% and 2.2% on\nin-domain and out-of-domain distribution data, respectively. Further, we\nempirically show that our approach is more generic and adaptable to any SSL\nmethods and does not impose any additional overhead complexity. Besides, we\nalso outline the role of patch-based versus slide-based curriculum learning in\nhistopathology to provide practical insights into the success of curriculum\nbased fine-tuning of SSL methods. Code is released at\nhttps://github.com/srinidhiPY/ICCV-CDPATH2021-ID-8\n