Crynodeb
Detecting and classifying abnormalities in mammography presents a significant challenge, particularly during the early stages of cancer development. The primary objective of this thesis is to explore whether mammographic images canoffer deeper insights into the morphological/size progression of abnormalities. To achieve this, a range of deep learning techniques was applied, leveraging both bounding box-segmented lesions and full-view mammograms of in situ and invasive ductal carcinoma breast cancer.
Deep-learning classification models were integrated with dimensionality reduction techniques to identify and distinguish morphological and size-related characteristics through the analysis of clusters of deep features. Cluster separation was further examined using localisation and filtering methods, incorporating additional metadata associated with lesion morphology and size. This analysis demonstrated that the models organise data hierarchically, forming sub-clusters of lesions with shared characteristics. Moreover, when viewed from a global perspective, some clusters appear to exhibit progression patterns based on specific relevant features, although further investigation is required to substantiate this observation.
These results were further enhanced through the use of multi-view classification networks. Novel channel-wise attention mechanisms were introduced, providing a unique approach to integrating features from models analysing different viewpoints by extending squeeze-and-excitation blocks. One of the proposed channelwise attention mechanisms achieved performance comparable to a ResNet50 with squeeze-and-excitation blocks, with no observed loss in accuracy.
Additionally, the thesis includes a brief investigation into predicting breast cancer grade using mammography. For this task, several novel convolutional networks based on the rate of change were proposed. This collection of models incorporated multiple viewpoints, pre-screened samples, and the time interval between screenings. The investigation revealed challenges in accurately classifying different grade groups. Variations in the time between screenings and inconsistencies in lesion clarity led to erratic model performance, highlighting the inability of the methods to correctly distinguish between the different breast lesion grades.
Furthermore, an initial experiment was conducted to develop benchmarking data and deep learning methods for breast ultrasound. This modality can be more effective than mammography in certain cases, particularly for patients with dense breast tissue. This effort resulted in the creation of a reproducible benchmark dataset by consolidating multiple datasets and deep learning approaches, making it available to the broader research community. Nine state-of-the-art deep learning models were evaluated, including semantic segmentation models, fully connected U-Net variants, and transformer-based architectures. Mask R-CNN achieved the highest overall performance, with mean scores of 0.851 for Dice coefficient, 0.786 for intersection over union, and 0.975 for pixel accuracy. Statistical analysis using MANOVA/ ANOVA and Tukey tests confirmed that Mask R-CNN performed significantly better than all other benchmarked models.
| Dyddiad Dyfarnu | 2025 |
|---|---|
| Iaith wreiddiol | Saesneg |
| Sefydliad Dyfarnu |
|
| Goruchwyliwr | Reyer Zwiggelaar (Goruchwylydd) & Bernie Tiddeman (Goruchwylydd) |
Dyfynnu hyn
- Standard