Open Collections

Wiebe, Mitchell

University of British Columbia

Purpose: Histologic evaluation of biopsied lesions is required for definitive diagnosis of lung cancer. Deep learning has been shown capable of pathologist-level diagnosis of lung biopsies at high spatial resolutions (0.25-2μm/pixel); moreover, lung histology has been visualized with tomographic imaging at 7.4μm/pixel. This work investigates the effect of spatial resolution (SR) on differential diagnosis of lung histology images below 2μm/pixel, which, to the best of our knowledge, has not yet be investigated. Methods: Pathology slides from The Cancer Genome Atlas (adenocarcinoma (LUAD) = 823, squamous-cell carcinoma (LUSC) = 753, and normal = 591) were used to train (70%) and test (30%) a convolutional neural network (Inception-v3) as a binary (cancer vs. normal) and three-way classifier (LUAD vs. LUSC vs. normal). Slides were partitioned into tiles at magnification levels corresponding to SRs of 4, 8, 16, 32, and 64μm/pixel. Additionally, slides were tiled into 512×512 pixel tiles at 2.5× magnification (SR = 4μm/pixel) and reduced SRs of 8, 16, 32, 64, and 128μm/pixel were simulated with Lanczos3 low-pass filters. Slide-level predictions were obtained by averaging constituent tile predictions and performance was evaluated by area under the ROC curve (AUC). Confidence intervals (95%) were determined by bootstrapping and an arbitrary performance threshold was set at AUC = 0.95. Results: At low magnifications, tiled datasets were small, and led to model over-fitting or poor performance due to lack of representative histology patterns. This problem was overcome by using a larger dataset for model training (2.5× = 4μm/pixel, 512×512) and simulating lower SRs with Lanczos3 low-pass filters. For the binary classifier, the minimum SR that was classified within tolerance of the performance threshold was 64μm/pixel (AUC = 0.980, CI = 0.963-0.992). For the three-way classifier, the minimum SR that was classified within tolerance of the performance threshold was 16μm/pixel (AUC-LUAD = 0.940, CI = 0.920-0.957), (AUC-LUSC = 0.940, CI = 0.922- 0.957), and (AUC-Normal = 0.992, CI = 0.984- 0.998). Conclusion: Deep learning can be used to differentiate cancer vs. normal lung pathology at SR = 64μm/pixel and LUAD vs. LUSC vs. Normal at SR = 16μm/pixel.

Text

2022-08-19

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/82419

Medical Physics

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/