UC Davis

The biomedical field is being revolutionized by deep learning. The accumulation of more, and more complex data opens the opportunity for the development of new advanced deep learning models that can address fundamentally difficult question in biology. Nevertheless, the integration of deep learning into the life sciences is complicated and necessitates a careful strategy in order to tackle specific domain challenges while optimizing the vast quantities of very heterogeneous data. This thesis addresses very different biological problems using deep learning methods. A common theme to these problems is the complexity and richness of the large data sets collected.In chapter 2, I introduce \textit{Scratch-AID}, an exceptionally precise deep learning-based system designed to automate the identification of mouse scratching in a controlled environment. By employing a convolution recurrent neural network trained on video data, this system attains a recall percentage of 97.6\% and a precision percentage of 96.9\%. Therefore, \textit{Scratch-AID} can be considered a feasible substitute for manual quantification techniques in the context of pharmacological screenings and behavioral studies.

Following this, in Chapter 3, I address the problem of survival in cancer patients. I present a deep learning framework called \textit{Genomic-Guided Hierarchical Attention Multi-Scale Pathology Transformer} (GG-HAMPT). I introduce an multi-level early data fusion strategy that uses the co-attention paradigm to combine genomic data and 3 level representation of whole slide Images (WSI) of cancer tissues. Employing the co-attention module, our model adeptly discerns the significance of pathological patches in correlation with grouped genomic features. In the context of predicting survival outcomes using multimodal data, including gigapixel WSIs, our GG-HAMPT model outperformed existing other weakly supervised approaches.

Last, we examine the domain adaptation of pre-trained models for the purpose of classifying digital pathology images through the implementation of the LoRA Vision Transformer (LoRA-ViT). The model exhibits remarkable efficiency and accuracy. Furthermore, in a 6-class cell classification task, LoRA-ViT outperformed the native ResNet 50 model. In experiments conducted with constrained datasets, the LoRA-ViT model significantly outperformed the ResNet 50, showing its superior generalizability compared to the baseline model.