Investigating Bias and Model Efficiency in AI for Lung Ultrasound Classification – research by Srishti Jain and Saunak Bhattacharjee
This research was conducted by GRS alumni from BU’s Department of Computer Science.
Artificial Intelligence is rapidly advancing in healthcare, particularly in diagnostic imaging. However, as powerful as these models can be, their reliability hinges on the quality and characteristics of the data they’re trained on. Two MS in AI Alums, Srishti Jain (GRS ’25) and Saunak Bhattacharjee (GRS ’25), have been working at the intersection of AI and medical imaging to ensure that such models are both effective and trustworthy.
Their recent work focuses on lung ultrasound video classification, with an emphasis on bias detection and model efficiency. Srishti and Saunak have been invited to present their research findings at the following conferences:
- CMIMI 2024 (Boston University) – Poster and oral presentations were delivered at the Clinical and Medical Imaging and Informatics conference held in October 2024.
- IEEE IUS 2025 (Utrecht, Netherlands) – Their work will also be featured at the IEEE International Ultrasonics Symposium in September 2025.
Below is an overview of their contributions through two research papers.
Paper 1: Unveiling Bias in AI Model Training Data
In this study, Srishti and Saunak explored an important question: “Can AI models unintentionally learn biases from the structure of medical imaging data?”
Specifically, they looked at lung ultrasound videos used for diagnosing pneumonia. In clinical practice, ultrasound technicians often vary their scanning approach, moving the probe at different speeds or spending more time in certain areas of the lung. As a result, the videos collected for training AI models can vary significantly in length and content.
The concern was that an AI model might pick up on non-clinical cues, such as video length, rather than meaningful medical features. For example, a model might infer that longer videos correspond to pneumonia cases, which could lead to inaccurate predictions in real-world settings.
To investigate, they ran controlled experiments to test how much influence these intrinsic variations had on model performance. Their results were promising: the model they trained was robust and did not appear to rely on such biases. But the study makes an important point: bias audits like this one are necessary before implementing AI models in the healthcare industry, where misclassification can have major consequences.
Paper 2: Comparing Classic to State-of-the-Art Image Features
The second paper presents a comparative study of two different feature extraction methods used for classifying lung ultrasound videos:
- Local Binary Patterns (LBP) – a traditional, low-complexity method
- ResNet-18 – a widely used deep learning architecture
While ResNet-18 is known for its accuracy and complexity, it also requires more computational resources. Surprisingly, the results showed that LBP performed nearly as well as ResNet-18 in classifying pneumonia cases. This finding suggests that lightweight models and classical techniques may still hold significant value, especially in settings with limited computational infrastructure.
This comparison is not only important from a performance standpoint but also opens up possibilities for building efficient, cost-effective diagnostic tools in resource-constrained environments.
The study conducted by Srishti and Saunak adds to the expanding corpus of research on reliable, explainable AI in medicine. Their focus on bias analysis and effective model design is especially relevant to clinical AI applications where reliability and transparency are crucial. Additionally, by enabling AI models to learn from minimally structured information, their work expands on self-supervised and unsupervised learning approaches. This approach provides guarantee for increasing AI applications in medicine across a variety of real-world contexts. As AI develops, studies like these remind us that model accuracy is only one aspect of the equation: equity, efficiency, and clarity are just as important.
Big kudos to Srishti and Saunak for not just pushing boundaries—but doing it responsibly.
To learn more about our master’s programs at the Department of Computer Science, check our page here.
About The Author:
Harini Saravanan
Marketing & Communication Assistant, GRS
Master of Science, Applied Data Analytics