Recent years have seen an exponential increase of deep learning applications throughout multiple fields including transport, entertainment and healthcare to name but a few. The growth in deep learning applications has been driven by more accessible computational power and the constant stream of new data that has enabled machine learning become an important feature of our daily lives.
My name is Gabriele Aldeghi and I recently joined the rapidly expanding ValitaCell AI team. I thought it would be interesting to reflect on, and share with you, my career journey to date. My journey to learn more about computer science and machine learning led me to the US, where I worked with the Ophthalmology department of the University of Illinois at Chicago Hospital.
The challenge
In the US, I collaborated in close contact with a team of ophthalmologists where we focused on a surgical procedure for removing Epiretinal Membranes, commonly found in the elderly.
These fibro cellular membranes develop inside the eyeball, over the inner surface of the retina. They can impact the eyesight of the patient and the severity can range from visual distortions to severely impaired vision.
The surgical operation to remove this membrane is performed by accessing the eyeball via small incisions. Using surgical instruments in such conditions is very demanding; the restricted movement of the instruments and the inherent fragility of the retina pose challenges to ophthalmologists and risks to patients.
The Deep Learning Opportunity
It was evident that deep learning could play a transformational role in these complex surgeries by providing visual cues to surgeons. To support surgeons, intraoperative Optical Coherence Tomography (iOCT) devices are used to supply real-time cross-section OCT images of the region of interest. These images provide important cues about tissue alteration during the surgery, thus helping the surgeon avoid any accidental retinal damage that would impair the patient’s eyesight.
Unfortunately, to provide real-time images, the iOCT cannot provide the same image quality as OCT. The quality reduction affects the ability of surgeons to identify tissue alterations. For this reason, having an algorithm to identify these tissue alterations automatically would be beneficial for the surgeon.
My thesis at the University of Illinois at Chicago and Politecnico di Milano focused on developing a deep learning algorithm with state-of-the-art techniques to identify these tissue alterations. The goal was to predict a binary semantic segmentation: given an input image, I wanted to identify where the retina layer was in the image.
I leveraged UNets, different regularization techniques, and heavy specialized data augmentation pipelines to obtain results close to expert-level performances. Moreover, the proposed deep learning model was able to produce real-time images that would be able to support ophthalmologists during surgery. The application of these deep learning initiatives can provide a significant level of support for skilled surgeons during these complex procedures. It is clear that healthcare has much to benefit from the application of deep learning.
CellAi®
Today, I am working with the team at ValitaCell to pioneer new analytical technologies for the biopharmaceutical industry. I am applying my expertise to ValitaCell’s Artificial Intelligence (AI) initiatives including CellAi® which is our Ai powered, image analysis software. I am creating deep learning models and currently working on predicting the viability of Chinese Ovary Hamster cells via brightfield images. This problem involves using Convolution Neural Networks to process the images and using Explainable Artificial Intelligence to understand how these complex networks work underneath.
Using machine learning to predict viability measurements is a non-invasive way, stain-free approach and is a transformational approach to extracting rich cell insights in a non-destructive manner. I’m looking forward to being part of ValitaCell’s exciting AI journey. You can also check out my colleague’s blog post about label-free imaging and its advantages here.
