GISR-Net: An image-to-scalar regressor for quality assessment of illumination correction in dermatological macro-photographs

Owing to their simplicity and flexibility, macroscopic photographs captured with ordinary digital cameras are commonly employed in dermatology as an alternative to conventional dermoscopy images. Macroscopic photographs often suffer from uneven background illumination, which degrades segmentation accuracy and diagnostic decisions in automated systems. Illumination correction is often used as a preprocessing step to address this issue. No-reference image quality assessment (NR-IQA) metrics are therefore necessary to evaluate the quality of macroscopic photographs at acquisition and after illumination correction. Such image quality assessment (IQA) metrics can also serve as a cost function for optimizing user-defined parameters of illumination correction algorithms. NR-IQA metrics whose quality ratings agree well with segmentation accuracy are not available. This paper proposes a task-based deep learning model, the GoogLeNet-based image-to-scalar regressor network (GISR-Net), for assessing the quality of illumination-corrected macroscopic photographs. GISR-Net is derived from GoogLeNet by modifying the final fully connected layer to produce a scalar output instead of class labels and by replacing the softmax and classification layers with a regression layer. Transfer learning was used to reduce the dataset size and training time required. To train GISR-Net, illumination-corrected images of different quality levels, generated by 16 state-of-the-art algorithms, are used as inputs, and the corresponding mean segmentation accuracy obtained with Otsu’s, minimum error threshold and Kapur’s maximum entropy segmentation is used as the target. The agreement of the proposed IQA metric with segmentation accuracy is evaluated using Pearson linear correlation coefficient (PLCC), Spearman’s rank-order correlation coefficient (SROCC), and Kendall’s rank-order correlation coefficient (KROCC). The proposed IQA metric (PLCC = 0.66, SROCC = 0.67, KROCC = 0.50) correlates more strongly with segmentation accuracy than lightness-order error (PLCC = 0.10, SROCC = 0.09, and KROCC = 0.06) and the average gradient of the illumination component ( PLCC = 0.35, SROCC = 0.08, and KROCC = 0.05). The proposed task-based NR-IQA metric measures the quality of illumination-corrected dermatological macroscopic photographs and correlates with segmentation accuracy.