Interactive Tumor Segmentation and Analysis Toolkit
Building an interactive web-based, and transformer-backed automatic segmentation toolkit for quick lesion segmentation from 3D inputs like CT Scans. Text-based input is also supported for better accuracy.
From tabular data output from multi-channel images, construct graphs to capture relative information among nodes and subsequent classfication via unsupervised learning.
Built an interactive visulziation tool for high-dimensional embedding interpretation. The embeddings were usually generated as unsupervised outputs and the tool can cluster (HDBScan, UMAP), view (heatmap, histogram, archetype) the results and thus help the users to intepretae unsupervised learning outcomes.
Unsupervised 3D Image Analysis with Patch-Based Approach
Processed large amount of 3D medical images using a patch-based approach (otherwise the entire images are too large to fit into GPU), followed by unsupervised learning and embedding analysis.
3D deep learning model training for chest vessel segmentation and atherosclerosis identification using MRI data.
Integrated Tumor Segmentation Solution Using MRI Scan as Input
Dockerized solution for tumor segmentation in lung cancer with 3D bounding box visualization output.
Lung Cancer Prediction Based on Patient Metadata
Lung cancer and subtype prediction using XGBoost based on metadata from 1658 patients.
Automatic Identification of Rhythm/Morphology Abnormalities in 12-Lead ECGs
Model training and network tuning (ResNet variant) to identify abnormal signals in 12-lead ECG data from 7191 samples.
Colorectal Surgical Phase Detection Based on Endoscope Video Records
Prediction of colorectal surgical phases using video and sensor data.
As a teaching assistant, I coached students these projects. (2018 - 2019)
Automatic System Providing Matching Images for A Given Article (NLP)
An automatic system finding matching images for a given piece of text (articles, news, URL, etc.). First generating keywords for a given article written in natural language; then returning images with matching keywords. Image keywords are generated by deep learning.
Information Highlighting Based on Impact from A Given Text (NLP)
For a given article or URL, highlight content causing positive and negative impacts on a certain topic in different colors.
Identification of Genomic Sites Affecting Oil Production in Soybeans
Using single-nucleotide polymorphism (SNP) information as features to predict oil production in soybean strains with gradient boosting methods (XGBoost, LightGBM, etc.).
Tumor Segmentation in Lung Cancer
Train and tune a deep learning model to segment suspicious regions in potential lung cancer cases using medical MRI data.
As a student : )
Microcalsification Segmentation Using Deep Learning
Deep learning model training and prediction (using U-Net architecture) to identify microcalcification from mammography
Identifying Novel Genes Using Ribosomal Profiling and RNA-Seq Data
Potential genes discovery by investigating previously unannotated regions in genome using ribosomal profiling and RNA-Seq data.
Differential Expression Analysis
Alignment, mapping to discover function of introns and revolutionary significance in several species from fungus to high mammal through bioinformatic methods.
As someone who enjoys coding and trying new techs.
Music Gen
Latest project. A decoder-only generative transformer including basic infrastructures. Writen with PyTorch and supports both text and midi files as input.
For text inputs it’s a language model (just like LLM but much smaller in terms of blocks and heads); and for MIDI inputs it can generate music. They use different tokenizers of course.
An emulation of Enigma machine M3 implemented in Python 3.
After reading about this type of encryption invented almost one century ago I was fascinated about it; as well as the story that our AI patriarch Dr. Turing decoded those encrypted messages.
So I decided to realize it with codes. It was pretty object-oriented with the machine and components (rotors and reflectors) realized in classes.
Rendering method that fakes a 3D space on screen by using only a 2D array, achieved with geometric tricks. The game Wolfenstein 3D (1992) used this rendering method.
An experiment for video cutscene and programmed object movement in Unreal Engine 4.22.
Physics engine enabled, raytracing enabled. Made with free assets.
Soundboard App with Flutter
A sound board app with audio mixing, shuffling, video playback and motion control functions. Initially written in Swift 5.0 and later re-written in a multi-platform framework (Flutter + Dart).
Originally a prank but later on became an awfully fun project.
* All images used in the project are royalty free. Animations by Adobe After Effects Icon by Affinity Photo.
Re-implementation using Flutter. Showing the macOS build.
Windows 11 build. Also has Linux binaries enabled by GitHub Actions.
Online Audio Cutter
A web service enabling audio extraction for given time intervals from pre-defined videos. Previously submitted requests are shown on the right side and are available for download. I use this service to extract audio fragments for soundboard App mentioned above.
Automatic Identification of Rhythm/Morphology Abnormalities in 12-Lead ECGs
8 types of disease detection using 12-lead ECG data from 7191 samples
In the context of a deep learning competition Displaying first channel of input and corresponding weights of the first type extracted from the last layer
Chest Vessel Classification And Atherosclerosis Detection