Research Project
10325131
PROJECT SUMMARY For adequate diagnosis and staging, transurethral resection of bladder tumor (TURBT) specimens must extend into the bladder muscle wall. Studies indicate that for patients with high-grade bladder cancer, 5-year mortality was 8% when the muscle was present in the TURBT specimen, and 13% when absent. For this reason, if there is not sufficient muscle in the specimen after the initial resection, guidelines recommend repeat TURBT. Almost half of TURBTs do not contain muscle as confirmed post-operatively by histopathologic examination. There are currently no practical tools available to surgeons to determine during the procedure whether the resected specimen includes sufficient muscle tissue. The goal of this project is to develop an imaging device that will be used for point-of-surgery detection of muscle in TURBT specimen in real-time. We will use ultraviolet light-emitting diodes to selectively excite different native fluorescent molecules in the tissue sample. We will further increase the biochemical information content by complementing the autofluorescence data with multi-wavelength reflectance images. We hypothesize that the combined multi-spectral autofluorescence and reflectance images will provide a snapshot of the integral biomolecular information of the tissue and, when combined with deep learning, capture latent biochemical and morphological differences that are encoded in the multispectral images. Our hypothesis is based on the fact that the connective tissue lamina propria and epithelial tissue have different biochemical make-up than the muscularis propria. We will employ a deep learning framework on the acquired images to develop a training algorithm from >200 ex vivo TURBT specimens from > 50 patients. The measured tissue will be processed for histopathological investigation to create true labels for algorithm training. We will interpret the deep learning classification results by correlating the extracted class features from the trained neural network with input image parameters, and consequently attribute them with known biological differences of the tissue types. To test the algorithm, we will acquire independent image sets from 80 samples from 20 patients and assess the concordance between our results and pathologists? reading of the Hematoxylin and Eosin (H&E) slides. We will also use a convolutional neural network trained using a generative adversarial-network model to transform wide-field autofluorescence images acquired from unlabeled tissue sections into H&E images of the same samples. The virtual H&E images will be evaluated by pathologists to recognize major histopathological features in images generated with our virtual staining technique and compared with the histologically stained images of the same samples.
