Portfolio of machine learning projects completed for academic & self-learning purposes.
It's better to look at it on this link instead of on github.
A project with 📜 means I had some external help (project as part of a course or help from a teacher).
The goal of the project is to be able to predict the rating of a review based on text review.
| Reviews Text | Reviews Rating | |
|---|---|---|
| 1 | This hotel was nice and quiet. Did not know, ... | 3 |
| 2 | Not cheap but excellent location. Price is ... | 5 |
| 3 | Parking was horrible, somebody ran into my ... | 3 |
| . . . | . . . | . . . |
After preprocessing the text (tokenization, stopwords, lemmatization, etc.), I used SentimentAnalyzer in the nltk library to rate the sentiment of a review (~ 1 means positive, ~ -1 means negative). From that grade I build a classifier that would predict the rating of the review.
Another method I implemented was to use fasttext, a library for learning of word embeddings and text classification created by Facebook's AI Research lab.
The goal of this project is to generate a synthesis of the data created by French citizens to answer some political open questions.
TSNE Visualization of corpus
My approach focused on the environmental questions. I used N_Grams and LDA to pick up the subjects that came up most often in the answers. The results were issues like overconsumption, greenhouse gas emissions, extinction of species, fossil/nuclear energy, etc. Finally, I looked at whether those subjects emerged in the summary of Emmanuel Macron's appearance in the Grand Débat emission.
This project was an introduction to the use of CNNs. Using Selenium, we scrapped images of both akita inu and giant schnauzer.
Example of images (giant schnauzer on the left, akita inu on the right)
PPG-DaLiA is a multimodal dataset featuring physiological and motion data, recorded from both a wrist- and a chest-worn device, of 15 subjects while performing a wide range of activities under close to real-life conditions. This dataset is designed for PPG-based heart rate estimation but our instructor challenged us to predict the activity of the subjects.
Plot of the heartrate during the different activities.
The purpose of this project was to deal with diverse sources of data and determine how to make use of them for our task. For example, each sensor had a different frequency so understanding and standardizing the different attributes was an essential part of this project.
Since the dataset provides the data for only 15 subjects, I aimed my study at looking how the model developed would perform on unseen data.
The goal of this project was to build a chest X-ray classifier based on the ChestX-ray8 dataset. Each of the X-ray images is associated with 14 different pathological conditions.
We used transfer learning to retrain a DenseNet model, built a weighted loss function to address class imbalance, and visualized our model's activity using Grad-CAM.
Example of the computed activity of an Xray image.
The goal of this project was to build a neural network to automatically segment tumor regions in brain, using MRI scans.
Visualization of an MRI image.
We generated sub-volumes of our data and built a 3D U-net network. We looked at different metrics to evaluate our model (soft dice loss, dice coefficient, sensitivity, specificity) and applied those on entire scans.
Comparison between prediction and ground truth.
The goal of the challenge is to detect lymph node metastases in histological images of patients diagnosed with breast cancer. Each patient is described by 1,000 small images (tiles) extracted from one whole-slide image.
Example of two tiles (non-tumoral on the left, tumoral on the right)
Additionally, we have 11 patients for which each tile was annotated by a pathologist (total of 10,024 annotated tiles). Based on these annotated patients, I built a model that predicts the probability of metastases in the tile. Next, I computed the probabilities for every tile of the non-annotated patients. Based on these predictions, I built a model that predicts whether or not a patient has any metastases in its slide.
The goal of the challenge is to predict survival time of patients diagnosed with lung cancer, based on 3-dimensional radiology images (CT scans). Clinical data of patients and radiomics (quantitative features extracted from the scan) are also provided.
CT Scan
Mask of tumor (green part)
Based on the imaging modality (scan & mask), I re-computed radiomics features using PyRadiomics (the radiomics data provided was only a subset of what could be computed using that library). After a feature selection process using VIF, I built a Cox Proportional Hazard model to predict the survival time of the patients.
The goal of this challenge is to segment liver lesions in contrast-enhanced abdominal CT scans. Liver segmentation and tumor burden estimation are also evaluated. The data and segmentations are provided by various clinical sites around the world.
Example of a CT scan
The first step of the process was to create a function that generates sub-volumes from our images. I also made sure that the sub-volumes extracted were relevant (i.e. at least 5% of something other than the background class). I am now building a U-Net architecture to complete the task of segmentation.