Hi there!

The MLA-LP model leverages the pre-trained CLIP vision-language model, enhanced with multi-level adapters and learnable prompts, to address the challenge of detecting and segmenting abnormalities in medical images. The model excels in both zero-shot and few-shot learning scenarios, achieving superior performance in anomaly classification (AC) and anomaly segmentation (AS) tasks across diverse medical imaging datasets. Key Features Multi-Level Adapters: Adapts intermediate layers of CLIP’s visual encoder to capture medical-specific features. Learnable Prompts: Dynamically adjusts textual prompts to improve flexibility and generalization across medical datasets. Zero-Shot and Few-Shot Learning: Performs effectively with minimal or no labeled data. Datasets Evaluated: ChestXray, HIS, OCT17, BrainMRI, LiverCT, RESC. Performance Few-Shot Setting: Achieves an average AUC improvement of 0.28% for anomaly classification and 0.08% for anomaly segmentation compared to state-of-the-art models. Zero-Shot Setting: Achieves an average AUC improvement of 0.39% for anomaly classification. Visualization Visualizations of anomaly detection and segmentation results are provided to illustrate the model’s capability to localize abnormalities in medical images.

Byte of Me is a personal website that works as both a portfolio and a content management tool. It supports multi-language UI with next-intl and uses a flexible database for dynamic content like blogs and projects. Built with Next.js, Supabase, and Auth.js, with a public site and a private dashboard for managing content.