Mingyuan Li

2023 Summer Undergraduate Research Fellowship (SURF)

Research Fellow

Supervisor:

Erick Purwanto

Replication and Improvement of the Cervical Spine Fracture Detection Research (Publish a paper at PlatCon)
• Implemented advanced pre-processing techniques, e.g., windowing, image cropping with Yolov5, and voxel clipping, to enhance image quality and prepare data for analysis.
• Developed a two-stage approach for cervical spine CT scan analysis, achieving a combined accuracy of 94.9%, and a BCE logits coefficient of 0.20 ± 0.01.
• Utilized UNet-EfficientNet in Stage 1 for precise CT image segmentation, attaining an accuracy of 99.91%. Applied CrackNet-LSTM in Stage 2 for accurate cervical spine fracture detection, achieving an accuracy of 94.9%.
Few-Shot Multi-Label Medical Image Classification Research (Publish a paper at CyberC)
• Training a few-shot classification model using the VPT model with a backbone of Swin-transformer. Evaluated the model's robustness using three datasets: ChestDR (chest X-ray), Endo (genuine colonoscopy), and Colon (colon cells).
• Developed multi-level attention patch-based preprocessing technique for the model, enhancing the model's ability to detect minute details, e.g., overlapping information.
• Improved mAP (1.2%-1.7%) and AUC (4.1%-5.2%) on two datasets (ChestDR, and Endo), compared to the baseline.

Teaching Development Fund Project: AI-based ChatBOT in Programming Education

Research Assistant (Team of Two)

Supervisor:

Erick Purwanto

Yihong Wang

Thomas Selig

Phase One: Created a programming tutor bot based on open-source projects tailored for programming and algorithm; winning the second prize at the XJTLU Student Research-Oriented Learning Summit
• Utilized Prompt Engineering to call OpenAI APIs, creating an educational chatbot for programming with 8 different topics. Implemented Chain of Thought and Moderation to prevent non-educational output, making it suitable for beginners.
• The bot was tested with an Indonesian student across four topics with two languages, achieving a 91% effectiveness rate, which demonstrates its pedagogical value and multilingual capabilities.
Phase Two: Developed algorithm visualization module based on bot responses to aid teaching; received the Global Impact Grants 2023-24 from Advance HE with present findings
• Created seven algorithm visualizations using d3.node, including for loops and array sorting lessons. Refactored the backend logic of a React-based TypeScript project to synchronize bot responses with dynamic visualizations, leading to the development of VisualCodeChat, a teaching-focused chatbot platform.
• Developed a pilot MOOC with VisualCodeChat and integrated our chatbot into an existing mature platform to create VisualCodeMOOC.
• Conducted 2 controlled experiments consecutively with 16 preA high school students and 88 non-programming background students from the CPT206 course. Reliable questionnaire results (Cronbach's alpha: 0.891) and a high average score of 4.18, supported by qualitative coding, demonstrate the bot met all evaluation criteria.
Phase Three: Researched the effectiveness and HCI criterias of our design compared to standard GPT in teaching graph algorithms.
• Iterated three graph algorithm visualizations using d3.force, including DFS-based cycle and connectivity checks, and BFS-based connected components check. Added data recording and user-friendly guidance modules.
• Conducted 8 controlled experiments to verify its effectiveness: experiments involved 330 students with varying programming backgrounds from the CPT204 course at XJTLU. One group used our bot for the set learning tasks, while the other group used AI based on the original GPT. The results showed a significant improvement in learning efficiency for students using the chatbot. The project was showcased at EdVenture and internal sharing seminars.