New Product Alert: ChexNet
With approximately 2 billion new lung disease cases worldwide annually, chest x-rays have become the most widely used technique in monitoring, diagnosing, and treating diseases such as cancer, pneumonia, and pleural thickening. While the reliability of the chest x-rays for imaging and diagnostics are improving, an estimated two-thirds of the global population has no access to diagnostics tools regarding radiology. To close this gap and make the health care services more accessible, we developed a deep learning-driven virtual advisory system that can detect up to 14 distinct thoracic pathology labels from frontal chest x-rays at an expert level.
Our solution aims to fasten the diagnosis procedure by leveraging deep neural networks’ ability to share standard low-level features that define lung tissue across all the disease labels and learn higher-level distinguishing features at the deeper layers. The developed model can be used as a consultation tool that can provide expert level diagnostic services in areas that lack a significant number of radiologists. Plus, it also can be integrated as a forecasting tool that can signal tendencies towards other diseases even though the patient necessarily has the disease at the current screening time.
Our model uses the datasets, collected with detailed research from contracted institutions and open source, containing 112,120 frontal chest x-rays from 30,805 unique patients indicating 14 distinct thoracic disease labels. The labels in these datasets are highly unbalanced, which creates a systematic bias in the learning process. Our solution overcomes this shortcoming by integrating custom activations to each label’s respective weight in the dataset as a post-processing step.
On the other hand, at the pre-processing stage, each chest x-ray sample is decomposed into five energy channels and iteratively normalized in these channels to a distinct mean value for each channel. This normalization aims to make the model agnostic to the source of the data. Different sources use different configurations on x-ray machines leading to contrast and brightness differences in the acquired images. The training is further augmented by using various data augmentation methods tailored precisely for medical data samples.
Accelerating and centralizing the radiological diagnostic pipeline paves the way for various economic benefits due to reduced time and effort by automating menial tasks. Using ChexNet, and similar technologies would decrease the costs and processing time of diagnosing lung diseases. With the ever-strengthening body of data, ChexNet would save health experts’ time and particularly help the doctors in less advantaged regions by allowing them to direct their attention to treatment rather than the pre-treatment stage. Invigorating healthcare applications with AI tools and better machine-learning-aware experts has two direct results: First and foremost, it benefits patients with faster and more accurate results. Second, it helps policymakers to build affordable and sustainable systems. And last, it functions as a go-to solution for doctors and health experts to fine-tune their results and reduce the friction caused by lengthy diagnosing procedures.
As Algomedicus, our sole purpose is to build agile health care products strengthened by the latest AI-applications. In the face of the most significant health crisis of the last century, every minute counts. The latest technologies should be utilized to help healthcare professionals save time and energy. Algomedicus is open to collaborating with all medical institutions to offer tailor-made AI-solutions for their ever-changing demands.