It all began with a simple idea fuelled by a deep passion. As a small business, we pride ourselves on personal attention and dedication to every detail. Our approach is rooted in quality and integrity, ensuring that everything we do reflects our commitment to excellence.

Description

Analyzing chest X-rays using machine learning algorithms is easier said than done. This is because the clinical labels required to train these algorithms are typically obtained through rule-based natural language processing or human annotations, which can introduce inconsistencies and errors. Additionally, it is challenging to collect datasets that represent a diverse range of cases and establish clinically meaningful and consistent labels using only images.

In an effort to advance the classification of X-ray images, Google researchers have developed artificial intelligence models to identify four outcomes in human chest X-rays: pneumothorax (collapse of the lungs), nodules and masses, fractures, and airspace opacification (filling of the lung tree with material).

In an article published in Nature , the team claims that the family of models, which was evaluated using thousands of images in datasets with high-quality labels, demonstrated "radiologist-level" performance in an independent review conducted by human experts.

The study was published a few months after Google AI and Northwestern Medicine scientists created a model that can detect lung cancer using screening tests better than human radiologists with an average of eight years of experience, and about a year after New York University used the Google Inception v3 machine learning model to detect lung cancer. AI also supports the tech giant's achievements in diagnosing diabetic retinopathy through eye scans, as well as DeepMind AI, a subsidiary of Alphabet, which can recommend the correct treatment for 50 eye diseases with an accuracy of 94%.

This new work used more than 600,000 images from two anonymized datasets, the first of which was developed in collaboration with Apollo hospitals and consists of X-ray images collected over many years from various locations. As for the second dataset, it is the publicly available ChestX-ray14 dataset released by the National Institutes of Health, which has historically served as a resource for AI efforts.

The researchers developed a text-based system for extracting labels using the radiological reports associated with each X-ray, which they then used to provide labels for more than 560,000 images from the Apollo Hospital dataset. To reduce errors associated with extracting text labels and provide appropriate labels for a subset of ChestX-ray14 images, they hired radiologists to review approximately 37,000 images in two buildings.

The next step was to create high-quality reference labels for model evaluation purposes. A panel process was adopted, in which three radiologists reviewed all final images of the settings and test sets and resolved disagreements through online discussion. According to the study's co-authors, this allowed them to identify and properly document complex findings that were initially detected by only one radiologist.

Google notes that while the models generally achieved expert-level accuracy, performance varied across the corpus. For example, the sensitivity of detecting pneumothorax among radiologists was approximately 79% for chest X-rays14, but only 52% for the same radiologists in another dataset.

The differences in performance between the datasets… highlight the need for standardized datasets of evaluation images with precise reference standards so that studies can be compared,” Google researcher Dr. David Steiner and Google Health technical lead Shravya Shetty wrote in a blog post. who contributed to the newspaper. “[The models] often identified results that radiologists consistently overlooked, and vice versa. Thus, strategies that combine the unique "skills" of both [AI] systems and human experts are likely to be the most promising for realizing the potential of AI applications in medical image interpretation."

The research team hopes to lay the foundation for more advanced methods using the approved label corpus for the ChestX-ray14 dataset, which they have made available in open source.

It contains 2412 images of training and validation sets and 1962 images of test sets, or a total of 4374 images.

During the study, two CNN models were created, and several trained models were used for transfer learning: VGG16, VGG19,
ResNet50, ResNet100, Xception, and InceptionV3. The obtained
training results of the models were analyzed, and options for increasing
accuracy were investigated. An example of integration in production is shown and possible
occurrence of errors at the stage of implementation in production is revealed.

If you have an interest in this neural network, and it can help you in the implementation of your business and other technical tasks, please send an application to email: info@ai4b.org