Abstract
Skin cancer is the most common cancer in the United States. Two photon excitation fluorescence (TPEF) and second harmonic imaging microscopy (SHIM) offer insights into skin cancer development without risk to the patient. Images acquired through these methods may be analyzed through U-Net, a deep learning architecture designed specifically for the task of image segmentation. The goal of this study is to lay the foundation for training a deep learning model capable of segmenting in vivo TPEF images. The study focused first on model training using images of cultured cells, followed by consideration of the feasibility of model training on in vivo images, acquired through TPEF imaging of SKH1 mice skin.
The first portion of the study was performed after building U-Net architecture through the Keras and TensorFlow libraries. Models were trained and evaluated on their ability to identify nuclear, cytoplasmic, or background regions. We studied how model performance was affected by training data class distribution, learning rate, and number of epochs. Additionally, the study considers which performance metric is most representative of model performance. Results from this study revealed a lack of meaningful dependence on class distribution. Ideal learning rates were found to be 7.5e-5, 1e-4, and 1.25e-4 due to increased output of acceptable models. Results surrounding the effects of epochs revealed that 100 epochs was most appropriate, though 75 epochs is also reasonable. Finally, the first portion revealed that the IoU metric was most closely associated with a model being deemed acceptable for use.
The second portion of the study was a preliminary investigation of the feasibility in training models using in vivo TPEF images. An acceptable model trained on in vivo TPEF images was not produced in this study, seemingly due to significant class imbalance and potentially insufficient training data. However, model training revealed the ability to reliably identify certain classes, reflected in individual class IoU values of up to 0.97. The best trained model yielded predictions with average IoU values ranging from 0.62 $\pm$ 0.17 to 0 $\pm$ 0. Further work for this portion remains, including acquiring greater training data, loss function adjustment, and class weighting.