Effect of contrast-limited adaptive histogram equalization on deep learning models for classifying bone scans.

Introduction: Contrast-limited adaptive histogram equalization (CLAHE) is one of the image processing methods to improve image contrast. We conducted this study to figure out how CLAHE affects the performance of deep-learning models for classifying bone scans.

Methods: One thousand seventy-two patients with bone scintigraphy were enrolled. These bone images were acquired 3 hours after injection. We used 654, 204, and 214 images as training, validation, and test datasets, respectively. All images were classified as normal or abnormal by the nuclear radiologist. The model for the experiment was based on VGG16 and some dense layers were added after the last layer of the VGG16 model for classification. We prepared the images without preprocessing, the images with histogram equalization (HE) applied, and the images with CLAHE applied. For each preprocessing method, training and testing of the models were repeated 15 times. The effect of preprocessing was evaluated as the mean AUC of each method.

Results: The mean AUC of images without preprocessing, with HE, and with CLAHE was 0.587 ± 0.026, 0.593 ± 0.043, and 0.649 ± 0.02, respectively. The images applied CLAHE as preprocessing showed higher performance rather than the images without preprocessing or images with HE (p < 0.001, p < 0.001). However, there was no statistically significant difference between images without preprocessing and images with HE (p = 0.633).

Conclusions: Applying CLAHE to images can help improve the performance of the bone scan classification model. When developing bone scan classification models, preprocessing that can handle appropriate contrast of the images such as CLAHE may be considered.