Browsing by Author "Gu, Qiangqiang"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Deep Learning in Digital Pathology
    (2023-07) Gu, Qiangqiang
    Digital pathology (DP), enabled by the availability of digitized whole slide images (WSIs), opens up possibilities for incorporating deep learning (DL) models into the development of computer-aided diagnostic (CAD) tools for cancer diagnostics. Among the various approaches, image classification and segmentation are widely utilized to enhance cancer diagnostics. Image classification provides slide-level predicted labels, such as tumor or non-tumor, while segmentation generates masks with x- and y- coordinates of predicted tumor areas. The scope of this dissertation research spans across multiple aspects. It involves the application of existing image classification models to differentiate between malignant breast cancer and normal breast WSIs. Additionally, a novel anomaly detection model was developed to identify anomalous tissues in melanoma WSIs. Furthermore, the developed anomaly detection model was effectively utilized for tumor segmentations in colorectal cancer (CRC). The contributions made by this doctoral dissertation research to the field of DP primarily stem from the development of the novel progressive context encoders anomaly detection (P-CEAD) model. This model successfully detects anomalies on melanoma WSIs and demonstrates extended applications for tumor segmentation on CRC WSIs. Furthermore, significant contributions arise from the utilization of existing image classifiers in differentiating malignant breast cancer from normal breast WSIs. The research findings shed light on the significance of hyperparameter configurations and dataset variations in the process of selecting model architectures. These findings highlight that non-specialized model architectures with optimized hyperparameter configurations, have the potential to surpass DP-specialized model architectures in achieving accurate classifications on binary breast cancer WSIs.