How PaddleX Enables Early Detection of Malignant Skin Tumors with AI Segmentation

Background and Motivation

In early 2024, the National Cancer Center reported a malignant skin tumor incidence of 2.4 per 100,000 in China, with roughly 35,000 new cases annually. Advanced melanoma has a five‑year survival rate of only 4.6% and a median survival of 1.42 years, highlighting the critical need for early diagnosis.

Industry Challenges

Dermatology faces high professional thresholds, uneven medical resource distribution, and a shortage of specialists, leading to delayed or missed diagnoses and increased patient burden. Imaging devices also require advanced image‑processing techniques to remain competitive.

Solution Overview

The Zhengzhou University Light Industry College Meikel Studio built an intelligent early‑screening system using the low‑code PaddleX platform. The system captures dermoscopic images, applies the PP‑LiteSeg semantic segmentation model, and automates lesion detection to reduce manual effort and misdiagnosis.

Data Collection and Preparation

To address data scarcity, the team aggregated the ISIC 2018‑2020 datasets, totaling 61,051 images from 2,056 patients covering seven disease categories. After filtering for quality, approximately 5,400 images (about 540 per class) were annotated using LabelMe and converted via PaddleX tools.

Data augmentation was performed with the EasyData platform to balance class distribution.

Model Selection

PaddleX offers two model families; benchmark results favored the lightweight PP‑LiteSeg‑T for its balance of accuracy and inference speed.

Zero‑Code Development and Data Validation

The zero‑code pipeline split and validated the dataset, producing visualizations of labeled samples and class distribution.

Model Training

After validation, the team trained the model with four GPUs, exposing key hyperparameters (learning rate, iteration count) in the front‑end for easy tuning.

Hyperparameter Optimization

Learning rate experiments identified 0.02 as optimal (fixed 1000 iterations). Subsequent iteration‑count experiments raised mIoU from 0.838 to 0.868, achieving a GPU inference time of 5.98 ms.

Deployment and Online Service

The model was deployed via the zero‑code pipeline, enabling API calls and a web UI for single‑image inference. Offline deployment packages are also available for edge devices.

Future Directions

PaddleX’s low‑code, private‑deployment capabilities allow the system to expand to other medical imaging tasks such as spinal NIFTI segmentation, and to integrate into edge devices for real‑time assistance, promoting broader AI adoption in healthcare and other industries.