Dr. Harry Pratt

Using mainly OpenCV, this analysis sought to determine if specific areas of the face are missed during routine sunscreen application and whether the provision of public health information is sufficient to improve coverage.

To investigate this, 57 participants were imaged with a UV sensitive camera before and after sunscreen application: first visit; minimal pre-instruction, second visit; provided with a public health information statement. Images were scored using a custom automated image analysis process designed to identify high UV reflectance areas, i.e., missed during sunscreen application, and analyzed for 5% significance.

Object detection and image segmentation revealed eyelid and periorbital regions to be disproportionately missed during routine sunscreen application. The provision of health information caused a significant improvement in coverage to eyelid areas in general; however, the medial canthal area was still frequently missed.

In this work and the consequent paper, I developed a Fourier Convolution Neural Network (FCNN) architecture whereby training is conducted entirely within the Fourier domain. The advantage offered is that there is a significant speedup in training time without loss of effectiveness.

Using the proposed approach larger images can therefore be processed within viable computation time. The evaluation was conducted using the benchmark Cifar10 and MNIST datasets, and a bespoke fundus retina image dataset.

The results demonstrate significant speedup without adversely affecting accuracy. This work was presented at the ECML conference in 2018.

The quantitative analysis of retinal blood vessels is important for managing vascular disease and tackling problems such as locating blood clots. Such tasks are hampered by the inability to trace back problems along vessels to the source accurately. This is due to the unresolved challenge of distinguishing automatically between vessel branchings and vessel crossings.

In this work, I presented a new technique for tackling this challenging problem by developing a convolutional neural network approach for first locating vessel junctions and classifying them as either branchings or crossings. We achieved a high accuracy of 94% for junction detection and 88% for classification.

Combined with work in segmentation, this method has the potential to facilitate automated localization of blood clots and other disease symptoms, leading to improved management of eye disease through aiding or replacing a clinician's diagnosis.

The diagnosis of diabetic retinopathy through color fundus images requires experienced clinicians to identify the presence and significance of many small features which, along with a complex grading system, makes this a difficult and time-consuming task.

In this work, I developed a CNN approach to diagnosis from digital fundus images that can accurately classify the disease severity. The developed model architecture alongside the image preprocessing and data augmentation can identify the intricate features involved in the classification task and their respective grading rules. The model can consequently automate the clinical screening process.

The model is trained using a GPU on a large publicly available dataset and demonstrates impressive results, particularly for a high-level multi-class classification task; it achieves a Kappa score of 0.75.

I created and put into production a CNN model that classified screenshots scraped from the app store in order to classify multiple features of a mobile game; including game mechanics, genre, theme, and more. The resulting classification vector and the extracted game names were compared with all other games using the resulting vector space and NLP.

This clustered all scraped games allowing us to track growing clusters and produce game similarity scores. The real-time calculation was highly optimizing using tensor operations in TensorFlow as there were thousands of new games scraped a day and hundreds of thousands of games to compare to in the resulting database.

Regardless, the calculations of the model and similarity check was able to be performed in milliseconds per game.

Through data analysis and clustering user cohorts with different lifetime values, behaviors and revenue were identified. This user event data was extracted from an SQL database, processed, and used to produce a machine learning model. The model predicted whether the user would enter a specific cluster in their interaction time to 90% accuracy.

This allowed personalized configurations for the user cohorts which optimized the amount of time they interacted with the game and their potential revenue.

The Python libraries metaflow and pandas were used to implement and put into production scripts that extracted data from an SQL database and generated a statistical analysis relating to A/B test metrics. The A/B tests related to ML-based model predictions would trigger the activation event if above a certain threshold. The analysis also produced statistical predictions for hypothetical A/B tests based on past data such as risked revenue, potential revenue uplift, and time required until a significant conclusion would be reached.

A U-Net-based CNN model was developed for a cloud-based API. The model produced a probability map of segmented features such as exudate and hemorrhages in retinal imaging which relate to eye disease. The physical area of these features was then quantified in order to determine the prevalence of the feature.

High dice coefficient and IoU metrics were achieved which generalized well across multiple ethnicities, nationalities, and camera types.

Both RCNN and YOLO CNNs were trained and put into production for a cloud-based image diagnosis API. The object detection model extracted multiple biomarkers within an image that were then segmented via a U-Net model in order to quantify the size and shape of the biomarker for medical analysis.

Rare diseases offer a unique challenge for deep learning models. Even the best data augmentation methods are often not sufficient enough to allow the classification of CNN models to generalize well to unseen data.

Therefore, this work utilized GANs in order to learn the latent space of rare diseases and produce more synthetic data from the learned latent space. These images were then used to improve the accuracy of the rare disease classification by an additional 20%.