Impact of Magnification, Image Type, and Number on Convolutional Neural Network Performance in Differentiating Canine Large Cell Lymphoma From Non-Lymphoma via Lymph Node Cytology.

Abstract

Background: Lymph node (LN) aspirates are often obtained to distinguish large-cell lymphoma (LCL) from non-lymphoma (NL) in dogs with enlarged lymph nodes.

Objective: Images from cytology slides tested the effects of magnification, image type, and number on a convolutional neural network (CNN) differentiating canine LCL from NL.

Methods: Three hundred images of LCL and NL were used to train a CNN and interrogate the effects of image magnification, type, and number on the model's performance. Identified cases were imaged at 200×, 500×, and 1000× magnification in color and gray-scale and then used to train and identify optimal magnification and image type. The impact of the image number per cohort (50, 100, 150, 200, 250, 300) on the top model's performance was then assessed.

Results: The highest performance with color images was achieved at 1000× magnification, with an accuracy of 95.8%, a Receiving Operating Characteristic (ROC) area of 0.997, and an F-measure of 0.958. Similarly, the best results with gray images, also at 1000× magnification, showed an accuracy of 96.67%, a ROC area of 0.994, and an F-measure of 0.967. Performance improvements were most significant and plateaued as the number of images per class approached 150, with an accuracy of 95%, ROC area of 0.939, and F-measure of 0.95.

Conclusion: The analysis across models suggests that color versus greyscale did not significantly impact overall performance to distinguish LCL or NL. Optimal magnification was 1000×. A minimum of 150 images per class is recommended for pilot CNN studies in this 2-class problem.