Intelligent Early Screening System for Malignant Skin Tumors Based on PaddleX Low‑Code AI

Early identification and treatment of malignant skin tumors are critical in medical technology. According to the National Cancer Center (Feb 2024), the incidence of malignant skin tumors in China has reached 2.4 per 100 000, with about 35 000 new cases each year. Advanced melanoma has a five‑year survival rate of only 4.6% and a median survival of 1.42 years, highlighting the urgency of early diagnosis.

The skin‑lesion detection field faces long‑standing barriers: high professional thresholds, uneven distribution of medical resources, patients often miss early lesions, a shortage of dermatologists, low screening efficiency, and high misdiagnosis rates. Manufacturers also need more advanced image‑processing and recognition technologies.

To address these issues, the Meikel Studio of Zhengzhou University of Light Industry developed an intelligent early‑screening system for malignant skin tumors using the PaddleX low‑code development platform. The system captures dermatoscopic images, applies the PP‑LiteSeg semantic‑segmentation model for precise lesion segmentation, and automates the workflow to reduce missed and false diagnoses.

Scenario challenges identified include: (1) hair and other noise obscuring dermoscopic images; (2) a large number of skin disease categories with imbalanced and diverse morphologies, requiring high‑quality data; (3) the need to protect patient privacy while ensuring high detection precision and fast inference; (4) difficulty in interdisciplinary collaboration between medical experts and algorithm engineers.

Data collection and preparation : The team collaborated with the Dermatology Department of Henan Provincial Hospital and obtained the ISIC2018‑2020 datasets, comprising 61 051 images from 2 056 patients covering seven disease types (e.g., actinic keratosis, basal cell carcinoma, melanoma, etc.). Approximately 540 images per class (total ≈ 4 200) were selected for annotation using LabelMe, then converted to PaddleX format.

Model selection : Two PaddleX model families were evaluated; the PP‑LiteSeg‑T model was chosen for its balance of accuracy and speed, as shown in the benchmark table (image omitted).

Training and optimization : Hyper‑parameter tuning identified an optimal learning rate of 0.02 with a fixed 1 000 iterations. Subsequent experiments explored the best iteration count, ultimately achieving a mean Intersection‑over‑Union (mIoU) of 0.868 (up from 0.838) and an inference latency of 5.98 ms on GPU. Training was performed on a 4‑GPU setup.

Zero‑code development : The PaddleX zero‑code pipeline provided data validation visualizations, automatic dataset splitting, and a user‑friendly interface for adjusting key hyper‑parameters before one‑click training.

Model deployment and demonstration : The trained model was deployed with a single click to an online service, exposing a RESTful API for real‑time image inference. Screenshots illustrate the deployment UI, API call method, and inference results. An offline deployment package is also available for edge devices.

Future outlook : The platform’s strong algorithmic and computational capabilities enable extension to other medical imaging tasks (e.g., spinal NIFTI segmentation) and to industries such as smart manufacturing, green energy, and precision medicine, fostering interdisciplinary innovation and sustainable technological advancement.