Projects

In problems of binary classification, an optimal model would have no false negatives or false positives. With current limitations to the training process, no model can reach 100% efficiency. Hence, prioritizing one case over the other proves to be our best performing model for real world implementation. In the healthcare field there is a large bias towards the minimization of false negatives. When detecting diseases, a false negative would mean the patient is at risk due to inaccurate diagnosis, and the hospital would not be as reliable. In comparison, a false positive would mean misused healthcare and economic loss. Pneumonia is one of the leading causes of death in nearly all age groups. False Negatives are extremely dangerous in comparison to the detection of pneumonia. To this end, we introduce a new method to reduce False Negatives when detecting pneumonia from a Chest Radiograph X-ray. Our proposed method involves the altering of data when retraining models to force it to have a bias of false negatives over false positives without losing the performance in comparison to the unbiased model. Our method improved our recall, the minimization of false negatives, by 20% with no change in the F1 score and 5% improvement in AUROC.

The segmentation of diseases is a popular topic explored by researchers in the field of machine learning. Brain tumors are extremely dangerous and require the utmost precision to segment for a successful surgery. Patients with tumors usually take 4 MRI scans, T1, T1gd, T2, and FLAIR, which are then sent to radiologists to segment and analyze for possible future surgery. To create a second segmentation, it would be beneficial to both radiologists and patients in being more confident in their conclusions. We propose using a method performed by radiologists called image segmentation and applying it to machine learning models to prove a better segmentation. Using Mask R-CNN, its ResNet backbone being pre-trained on the RSNA pneumonia detection challenge dataset, we can train a model on the Brats2020 Brain Tumor dataset. Center for Biomedical Image Computing & Analytics provides MRI data on patients with and without brain tumors and the corresponding segmentations. We can see how well the method of image subtraction works by comparing it to models without image subtraction through DICE coefficient (F1 score), recall, and precision on the untouched test set. Our model performed with a DICE coefficient of 0.75 in comparison to 0.69 without image subtraction. To further emphasize the usefulness of image subtraction, we compare our final model to current state-of-the-art models to segment tumors from MRI scans.

In modern times, Artificial Intelligence is becoming a very promising tool to solve many complicated problems such as autonomous cars. My project is to create an algorithm which will be able to recognize the people in the image as well as their emotions. This project lands in the category of Intelligent Technology because I am using Machine Learning to help the visually impaired people feel more connected to their loved ones. Machine learning is where the computer recognizes or does a certain task using its own inference and learning based on the input and output data. Machine learning is often recognized as a category of AI. AI stands for artificial intelligence which is simply a form of intelligence from computers or machines. We use AI in many different tasks, such as security, websites, face recognition, and more. The visually impaired people can’t see the people in front of them so may not know, how are they feeling. I want to help them feel more connected to their family and friends, and join in the conversation via understanding the emotions of people talking to them. To do this I am using machine learning and artificial intelligence to find and recognize people and their emotion in an image. The output will be spoken out of a small device.

Automatic Chest Radiograph X-ray (CXR) interpretation by machines is an important research topic. Pneumonia, a deadly disease, is diagnosed through CXRs and machine learning can accelerate this process. To this end, we present PneumoXttention, an algorithm that can detect pneumonia from a CXR image to compensate for human fallibility. The algorithm's architecture consists of an ensemble of two 13-layer convolutional neural networks trained on a dataset provided by the Radiological Society of North America, RSNA, containing 26,684 frontal X-ray images split into the categories of pneumonia and no pneumonia annotated by professional radiologists in North America. We validate PneumoXttention with impressive F1 scores on the test set, and against human radiologists on images drawn from RSNA and NIH, and also analyze PneumoXttention's usefulness in practice.

Thousands of people die due to heart attacks every year. Heart attacks are caused by the blockage in the arteries from calcium; this event is known as calcification. To quantify the amount of calcification in a patient’s heart, doctors and radiologists often use the gold standard exam of gated CT scans. Gated CTs are only taken when a patient is already suspected of having a heart problem. On the contrary, non-gated CT scans are more routinely taken because it is used for a larger spectrum of health issues. Given that non-gated CTs are more commonly taken, our U-net model’s purpose was to accurately predict risk on non-gated scans. While the model was trained on the gated scans of the Coronary Calcium and Chest CT dataset, we used it on the non-gated scans. We were able to derive a set of mathematical equations to automatically crop the non-gated scans to closely replicate the format of the gated scans. The performance of the model improved by 91% in mean absolute error (cropped scans - 62.38, noncropped scans - 674.19) and 32% in F1 score (cropped scans - 0.68, noncropped scans - 0.58)