Author ORCID Identifier
Year of Publication
Master of Electrical Engineering (MEE)
Luis G. Sanchez Giraldo
The emergence of deep learning models and their success in visual object recognition have fueled the medical imaging community's interest in integrating these algorithms to improve medical diagnosis. However, natural images, which have been the main focus of deep learning models and mammograms, exhibit fundamental differences. First, breast tissue abnormalities are often smaller than salient objects in natural images. Second, breast images have significantly higher resolutions but are generally heavily downsampled to fit these images to deep learning models. Models that handle high-resolution mammograms require many exams and complex architectures. Additionally, spatially resizing mammograms leads to losing discriminative details essential for diagnosis. To address this limitation, we develop an approach to exploit the relative importance of pixels in mammograms by conducting non-uniform sampling. More specifically, in this project, we combine the methodology proposed by Shen et al. for training a breast cancer classifier with the non-uniform sampling approach proposed by Recasens et al. On the CBIS-DDSM dataset, our method achieves an AUC of 0.8543 on the test set using input images of size (1152 x 896) and a custom partition and an AUC of 0.7819 on the test set using input images of size (576 x 448) and the official partition. Those results are superior to the performance achieved by Shen et al.: 0.8456 AUC using a custom partition and 0.7621 AUC using the official partition. The model performance demonstrates that non-uniformly sampled images preserve discriminant features requiring lower resolutions to outperform their uniformly sampled counterparts. We also show that the proposed method can be transferred to INbreast images without reliance on pixel-level annotations and boost the model performance on independent data.
Digital Object Identifier (DOI)
Posso, Santiago, "Nonuniform Sampling-based Breast Cancer Classification" (2024). Theses and Dissertations--Electrical and Computer Engineering. 198.