Alex Le

Debiasing Melanoma Images

Developed a deep learning computer vision program that generates melanoma images for people with pigmented skin tones with the goal of debiasing lighter skin predominant medical imaging databases.

Github Repo

Melanoma, a form of skin cancer that arises from melanocytes, forms 75% of skin cancer-related deaths in the US (Davis et al., 2019). Nevertheless, early diagnosis and classification of melanomas require clinical, dermoscopic, and molecular data and have challenges due to skin pigmentation differences in the patients, making it a growing field in Dermatology (Darmawan et al., 2019). According to Wolf et al (2013), machine learning models that are designed to tackle the melanoma detection problem remain highly inaccurate and yield approximately 30% incorrect predictions (Darmawan et al., 2019; Wolf et al., 2013). This has been causing inaccurate results, false trust in patients in cancer detection software, and even delayed melanoma detection (Wolf et al., 2013).

The use of deep learning models, such as convolutional neural networks, is found to be more accurate in identifying such cancers. However, since the melanoma prevalence in Caucasian backgrounds is at least 50% greater, the majority of these deep learning models are trained on data that is skewed toward lighter-skinned images (Raimondi, Suppa & Gandini, 2020). This underrepresentation in the machine learning models is causing high biased outcomes, which are unable to classify melanoma patients that have darker skin tones. (Guo et al., 2021).

Thus, I sought to utilize deep learning to correct deep learning model biases. By using a style transfer model architecture using convolutional neural networks (CNNs), I was able to combine databases of patients with darker skin and skin lesions from lighter-skinned patients to produce skin lesion images in pigmented skin that could be used for future deep learning model training to improve skin cancer detection in ethnic skin.

Result

Close Window

Maestro

Utilized a convolutional neural network to identify hand gestures and correlate it with music controls (ie. pause, resume, skip, reverse, etc.) with 99.58% recognition accuracy.

Github Repo

I wanted to create a gesture control system where different hand positions can activate different commands on a computer. I created a Convolutional Neural Network to recognize how many fingers (from 0 to 5) a right-hand holds up and maps the gestures to commands on a music playlist. This gesture recognition system allows for a different approach to controlling interfaces, increasing accessibility.

I wanted accurate recognition of input images so I tried to create a convolutional neural network for the task. I believe that introducing machine learning into the task would greatly benefit accuracy. Human hands are very complex because they can bend into various shapes and look different from different angles. This makes recognition difficult as different gestures may have similar contours. Nonetheless, I believe that remote control through hand gestures would offer an alternate approach to controlling interfaces, increasing accessibility for different groups of people. For those who may have disabilities or difficulties understanding or accessing traditional buttons, having an accurate and easily learned hand-gesture control would improve accessibility.

Preprocess input data to look like training data

Results

Close Window

Yoki

Developed a website that pairs registered users from a unique organization based on similar interests using a KD tree data structure to find the optimal compatibility between people.

Github Repo

Complex algorithm

We used a KDTree for the complex algorithm, in which each point in the tree represents a user's array of interest levels. Furthermore, we used the KDTree to create a sorted ArrayList of users with a special heuristic comparator our team designed. The heuristic comparator takes two users' interest level arrays and compares their values of the same interest categories one by one. We designed it so that the smaller the gap between two users' levels of interest, the higher their sorting priority. With this, we were able to create a list of "best" matches between a user and other users based on similar interests.

Since the KDTree's main purpose is to provide the best structure for quicker sorting on a fixed dataset, we programmed the KDTree so that it is rebuilt every time a user updates their interest or when a new user is added to the database. This is because the KDTree is created based on comparing an individual's interests for each depth/level in the tree. Changing a user's interests gives the user new "coordinates", thus the KDTree will no longer be the most efficient for the nearest neighbor searched. Thus, we reconstruct the tree every time a user's interest levels are modified. In addition, when a user does not specify their level of interest in a particular interest category, its default value will be 0.

Encryption

To ensure the safety and privacy protection of all of our users, we decided to implement encryption for user logs in and signs ups. We included a password encryption algorithm between the frontend and backend, as well as between the backend and database. This way, no outsiders can access the real value of the user's password when the different parts of our project communicate with one another. To create extra security, we made sure that we did not hardcode the cipher key in any of our project files. Instead, we have it stored in a separate key.txt file, which can only be accessed by a file reader. With the cipher key in a separate file, we will add it to .gitignore so that it will never be pushed to the repo and thus the public.

Close Window

Harvard Neuroscience Research

Analyzed rat behavior and its correlation with neurological activity. Constructed a convolutional neural network to identify unmarked 3D points using spatial and temporal data.

Github Repo

2021 REU Harvard System Biology Internship Sponsored by:

• Harvard Medical School
• Blavatnik Institute
• National Science Foundation
• Simons Foundation
• Harvard QBio
• Principal Investigator: Dr. Cengiz Pehlevan
• Primary Mentor: Ugne Kilbaite

From unlabeled 3D joint data, I was able to predict the skeletal structure of the rats and correlate it with neurological activity with 83% accuracy.

From predicted data, create a VAE that predicts the location of missing 3D data points

We do not know exactly how the nervous system controls motor behavior in health or disease. Many medical research studies to analyze neural activity, brain circuits, and biochemistry are initially conducted in rats. Currently, it is extraordinarily time-intensive for humans to manually observe, record, analyze, and interpret rat movements 24/7, making it difficult, if not impossible, to relate these behaviors to neural activity.

Therefore, I approached this problem using computational techniques. My project in the Pehlevan Lab was to develop a method for modeling the skeleton using a video feed of animal behavior. There were two major barriers to achieving this goal:

• We do not always know which 3D data point on the external body correlates to which internal joint
• Some data points may not be picked up by the cameras due to environmental noise.

Thus, I developed a machine-learning algorithm to predict the name of a 3D data point as well as predict the 3D coordinate of missing points. This algorithm consists of 3 major components: a neural network (NN) for the initial prediction of points, a graph neural network (GNN) for the temporal prediction of points, and a variational autoencoder (VAE) to predict the coordinates of missing points. Although there is existing software that can help determine the 3D points in space, the code is not optimized for rat behavior and does not account for missing points. Furthermore, we were unable to modify the code to fit different needs, since it is proprietary. Thus, my program developed an accessible tool that enabled the analysis of real-time rat behavior without time-intensive user input. We were in a good position to develop this program because we had access to large amounts of rat movement recordings and a robust computing system that could process this large dataset. As a consequence of my work, many labs will also be able to more powerfully and efficiently determine how mutations, neurological conditions, environmental exposures, or medical interventions affect rat movement and behavior, generating insights and strategies for improving human health.

Close Window

Genome Assembly

Takes a file of k-mers and infers an optimal sequence derived from a De Bruijn graph. The k-mers are first decontaminated and infrequent k-mers are corrected. The algorithm sucessfully obtained a 99.84% identity match with the SARS-CoV-2 genenome assembly.

Github Repo

The purpose of this program is to take in a file of k-mers and develop an optimal sequence derived from a De Bruijn graph. Before creating the De Bruijn graph, the k-mers are first decontaminated by known vectors, and then any infrequent sub-k-mers are then replaced with more frequent k-mers to ensure the integrity of the DNA samples. Decontamination is performed with BLAST-like seeding and extension. Once all the k-mers are decontaminated and corrected, the k-mers are placed in a De Bruijn algorithm that simplifies/condenses the graph to find the longest inferred sequence for the given k-mers. The program was able to take 50-mer reads and reconstruct 16 reads of length 1260 of the Sars-Covid-2 receptor binding domain.

In order to find the optimal parameters, I analyzed which values for k (k-mer length), t (threshold for frequent kmer), and d (hamming distance threshold). After analyzing the average ungapped global alignment score of each true/corrected read pair, I decided that these values optimal to reconstruct the genenome.

• 10 = kmer length to determine seeding in contamination
• 30 = kmer length of reads used to correct and construct a De Bruijn graph
• 17 = kmer length to determine seeding in correction
• 2 = threshold value for common kmer frequency in correction
• 2 = difference value for kmer variability by in correction

After the file of k-mers were decontaminated and corrected, the program created a De Bruijn graph that allowed me to infer the most likely sequence. The De Bruijn code also creates a simplified graph to reduce memory storage.

Example of simple De Bruijn graph compression

Example of complex De Bruijn graph compression

After applying the De Bruijn algorithm to the kmer reads, I was able to obtain 16 reads of length 1260.

Example of inferred reconstructed genenome

ATGTCTGATAATGGACCCCAAAATCAGCGAAATGCACCCCGCATTACGTTTGGTGGACCCTCAGATTCAACTGGCAGTAACCAGAATGGAGAACGCAGTGGGGCGCGATCAAAACAACGTCGGCCCCAAGGCTTACCCAATAATACTGCGTCTTGGTTCACCGCTCTCACTCAACATGGCAAGGAAGACCTTAAATTCCCTCGAGGACAAGGCGTTCCAATTAACACCAATAGCAGTCCAGATGACCAAATTGGCTACTACCGAAGAGCTACCAGACGAATTCGTGGTGGTGACGGTAAAATGAAAGATCTCAGTCCAAGATGGTATTGCTACTACCTAGGAACTGGGCCAGAAGCTGGACTTCCCTATGGTGCTAACAAAGACGGCATCATATGGGTTGCAACTGAGGGAGCCTTGAATACACCAAAAGATCACATTGGCACCCGCAATCCTGCTAACAATGCTGCAATCGTGCTACAACTTCCTCAAGGAACAACATTGCCAAAAGGCTTCTACGCAGAAGGGAGCAGAGGCGGCAGTCAAGCCTCTTCTCGTTCCTCATCACGTAGTCGCAACAGTTCAAGAAATTCAACTCCAGGCAGCAGTAGGGGAACTTCTCCTGCTAGAATGGCTGGCAATGGCGGTGATGCTGCTCTTGCTTTGCTGCTGCTTGACAGATTGAACCAGCTTGAGAGCAAAATGTCTGGTAAAGGCCAACAACAACAAGGCCAAACTGTCACTAAGAAATCTGCTGCTGAGGCTTCTAAGAAGCCTCGGCAAAAACGTACTGCCACTAAAGCATACAATGTAACACAAGCTTTCGGCAGACGTGGTCCAGAACAAACCCAAGGAAATTTTGGGGACCAGGAACTAATCAGACAAGGAACTGATTACAAACATTGGCCGCAAATTGCACAATTTGCCCCCAGCGCTTCAGCGTTCTTCGGAATGTCGCGCATTGGCATGGAAGTCACACCTTCGGGAACGTGGTTGACCTACACAGGTGCCATCAAATTGGATGACAAAGATCCAAATTTCAAAGATCAAGTCATTTTGCTGAATAAGCATATTGACGCATACAAAACATTCCCACCAACAGAGCCTAAAAAGGACAAAAAGAAGAAGGCTGATGAAACTCAAGCCTTACCGCAGAGACAGAAGAAACAGCAAACTGTGACTCTTCTTCCTGCTGCAGATTTGGATGATTTCTCCAAACAATTGCAACAATCCATGAGCAGTGCTGACTCAACTCAGGCCTAA

Close Window

RISD Industrial Design

Designed and built custom pair of shoes in the industrial design program at the Rhode Island School of Design for my final project.

My final shoe design is a cross between two traditional shoes: the Oxford and the Converse sneaker. This unique mix gives the shoe an old classic look with a modern-day twist. Furthermore, the pair of shoes is an inverse of each other to represent the duality of relationships and identity. To me, shoes have different purposes: running shoes, dress shoes, sneakers, racing flats, sandals, etc. The design was centered around the famous quote "walking in one's shoes", but many people carry various identities and use shoes for various aspects of their lives. Thus, by combining an Oxford and Converse as well as having inverse colors, these shoes represent the multifaceted identities of the user: old and new, traditional and modern. Each shoe also has a removable circular insert, which allows the user to customize their shoes with their own logo or personal emblem.

Shoe design build process.

First shoe I designed and manufactured.

Close Window

Google Biodesign

Awarded First Place in the inaugural 2021 Google Biodesign competition. Worked cross-functionally in collaboration with two RISD classmates to engineer a methodology to develop a biodegradable, eco-friendly printed circuit board using chitin from local shellfish

Google Biodesign First Place Winner
Close Window