Psoriasis, a prevalent inflammatory skin disease, affects 10-80% of patients with nail lesions, known as nail psoriasis. This condition causes cosmetic disfigurement, restricts daily activities, and diminishes quality of life, also serving as a risk factor for psoriatic arthritis. Early and effective therapeutic intervention, therefore, relies heavily on accurate and reproducible disease assessment. The Nail Psoriasis Severity Index (NAPSI) is a scoring system designed to quantify nail psoriasis severity, considering eight key nail findings (pitting, leukonychia, red spots in the lunula, crumbling, oil drop discoloration, onycholysis, nail bed hyperkeratosis, and splinter hemorrhage). NAPSI divides the nail into four quadrants and evaluates the presence of these findings in each, yielding a score from 0 to 8. However, the subjective nature of NAPSI criteria leads to considerable interobserver variability among dermatologists, hindering its widespread use. This study aimed to develop a deep learning-based tool to improve the reliability and ease of NAPSI scoring, thereby enhancing the accuracy and consistency of nail psoriasis assessment in clinical settings.

Several previous studies have explored the use of deep learning for NAPSI assessment. However, these studies present limitations. One study achieved high accuracy in recognizing individual nail psoriasis findings but required a specialized camera. Another study successfully segmented the nail into quadrants but lacked sufficient lesion detection accuracy. A third study used key point detection, but this approach may struggle with variation in hand shape and its accuracy hasn't been comprehensively validated. Finally, a study using multiple object detection models required manual image pre-processing. This current study aimed to create a user-friendly deep learning tool capable of accurate NAPSI calculation from standard clinical photographs, without the need for specialized equipment or image preparation. The goal is to provide a tool accessible to all clinicians, regardless of their dermatological expertise.

The "NAPSI calculator" was developed in two steps: nail detection and NAPSI scoring. **Step 1: Nail Detection:** This step used a Single Shot MultiBox Detector (SSD) with a VGG16 backbone pre-trained on ImageNet. 995 hand and foot images from Google Image Search were used for training and validation, with 881 images from 78 patients at Keio University Hospital serving as the test set. The model was trained for 500 epochs, and performance was evaluated using mean average precision (mAP), calculated ten times with newly split training/validation sets and reinitialized network parameters. **Step 2: NAPSI Calculation:** This step involved two separate VGG16 models, one each for nail matrix and nail bed assessment. These models classified images into five classes (0-4 points), and their scores were summed to compute NAPSI. The dataset for this step comprised 2939 fingernail images (excluding toes and low-resolution images) manually cropped from the images used in step 1, divided into training, validation, and test sets. Nine board-certified dermatologists annotated NAPSI for each image. A custom loss function, combining cross-entropy loss and the squared difference between calculated and annotated NAPSI, was used to enhance accuracy. The model was trained for 200 epochs, with performance evaluated ten times using micro and macro average accuracy, considering an error within one point as accurate. **Step 3: The "NAPSI Calculator":** This integrated the best-performing nail detection model from Step 1 and the ten best NAPSI calculation models from Step 2. The final test set included 29 hand and finger images (138 nails) from 12 new patients with nail psoriasis. NAPSI was annotated by the same nine experts. The "NAPSI calculator's" accuracy was compared to that of six non-board-certified residents and four board-certified dermatologists (without nail expertise) using Welch's t-test.

In the final test set, the "NAPSI calculator" achieved 99.3% accuracy in nail detection (137/138 nails). The average micro-accuracy for NAPSI scoring was 83.9% (95% CI, 81.6–86.2%), significantly higher than the 65.7% accuracy of six non-board-certified residents (p=0.008) and the 73.0% accuracy of four board-certified dermatologists without nail expertise (p=0.005). The "NAPSI calculator" consistently scored NAPSI within a one-point error range for each annotated score. The processing time for one hand image and one nail image was approximately 0.85 and 0.95 seconds, respectively, on a standard laptop. Grad-CAM analysis suggested that the model primarily focused on nail psoriasis findings, although some images indicated the model might not have fully learned all aspects of nail psoriasis features. The study also demonstrated the "NAPSI calculator's" ability to accurately track disease progression over time in a single patient.

The "NAPSI calculator" demonstrates significant promise in improving the reliability and efficiency of NAPSI scoring for nail psoriasis. Its superior accuracy compared to dermatologists with varying levels of expertise highlights the potential of deep learning to standardize and enhance clinical assessments. The tool's speed and ease of use make it readily adaptable to clinical workflows, potentially increasing the utilization of NAPSI and improving the consistency of care for patients with nail psoriasis. However, it's essential to address the limitations: errors exceeding two points were observed in some challenging cases, and the dataset may exhibit a bias towards severe cases, potentially affecting generalizability. Further studies with larger, more diverse datasets should be conducted to refine the model and validate its performance across different patient populations.

This study presents the "NAPSI calculator," a deep learning tool for reliable and accurate NAPSI scoring. The tool outperforms human scorers, is efficient, and user-friendly. Future work will focus on expanding the dataset diversity and improving the model's robustness to address limitations identified in this study. This technology has the potential to standardize nail psoriasis assessment and significantly improve patient care.

The study acknowledges several limitations: a potential bias towards severe cases in the dataset might limit the generalizability of the results; some challenging cases showed large discrepancies between calculated and annotated NAPSI; the Grad-CAM analysis suggests the model may not yet fully grasp all nuances of nail psoriasis features; and the datasets used for Step 2 and 3 are not publicly available due to consent restrictions.