Kewalin Rangsinaporn, MD, Radiologist and Director of Health Design Center at Bangkok Hospital Headquarters, opened the session with a thought-provoking statement: “While we may not be engineers, a basic understanding of AI is crucial before integrating it into our workflow.”
This resonates deeply with the radiology department, as we stand at the forefront of AI adoption due to our rich datasets, which fuel AI development. Driven by Bangkok Hospital’s commitment to excellent patient care through innovation, Kewalin, MD. shared their journey with AI in chest X-rays.
Why chest X-rays?
We have to admit that a plain chest x-ray is the most common medical performance for screening and diagnosis. However, the 2-dimensional nature of chest x-rays limits our visualization. With these two things stacked together, they contribute to human and systematic errors like workload fatigue, visual fatigue, decisional fatigue, inattentional blindness, and biases. These errors, in turn, can lead to malpractice and lawsuits.
Bangkok Hospital leads the way
Bangkok Hospital utilized a cloud-based AI system working seamlessly with its PACS system, transmits images directly to the cloud after deleting out any patient personal information, where AI algorithms analyze them in real-time.
“Building user understanding is crucial for successful AI integration in healthcare, and we need to build it for all users.” Kewalin, MD. added.
Perception and understanding of users
Bangkok Hospital utilized a cloud-based AI system working seamlessly with its PACS system, transmits images directly to the cloud after deleting out any patient personal information, where AI algorithms analyze them in real-time.
“Building user understanding is crucial for successful AI integration in healthcare, and we need to build it for all users.” Kewalin, MD. added.
- AI as a Supportive Tool:
AI is a supportive partner, not a replacement. Radiologists are supposed to read and analyze images independently first, then use AI as a second pair of eyes to confirm findings or highlight potential areas of concern. This collaborative approach fosters trust and ensures responsible AI use.
- Nothing is 100% Accuracy
There’s no such thing as a 100% accurate AI. Radiologists should not expect that AI always has the right answer and should understand how a particular AI algorithm was made to be able to understand its limitations.
- No One-Size-Fits-All AI
Just like doctors specialize, so too do AI systems. Each AI excels in specific areas. Some AI is for TB detection, some AI can detect up to 40 findings, and some can detect 8 findings.
- Feedback for Continuous Improvement
After we use AI, it’s important to provide constructive feedback to developers. This feedback allows developers to develop AI systems to learn and evolve, ultimately leading to more reliable and impactful tools for overall healthcare professionals.
The potential benefits of AI
The benefits of AI in radiology are not one-size-fits-all either. “The benefits you get depend on the pain points you face.” Kewalin, MD. added.
- Worklist Prioritization: While it is not a game-changer for private hospitals like Bangkok hospital where workload may be manageable, this feature becomes a vital lifeline for public hospitals where radiologists face a daily mountain of scans.
- Second Reader: Inattentional blindness and missed cases can appear anywhere even with the most skilled radiologists. This not only improves diagnostic accuracy but also minimizes the risk of missed findings, ultimately reducing legal battles.
- Auto-generated Reports: During the pandemic, the field hospitals were overwhelmed with patients and limited resources. AI’s ability to analyze X-ray images and generate preliminary reports. This feature is crucial for time saving.
Case demonstration
AI offers two main options of display: contours and heatmaps. It depends on your needs and preferences. However, the heatmap makes it easy to spot potential abnormalities at a glance. Kewalin, MD. shares with us plenty of cases, here are five examples that highlight the benefits of each approach:
Case 1: Comparison of 2 AIs from Eastern and Western countries
An AI system identifies an abnormality around the hilum and suggests “sarcoidosis” as the diagnosis. However, the radiologist recognizes this as a “tuberculosis,” a more common finding in Thailand compared to sarcoidosis. Radiologists must be aware of such limitations and exercise critical judgment when interpreting AI outputs.
Case 2: A pulmonary nodule in LUL
A pulmonary nodule in the left upper lobe sits near the clavicle, a challenging location due to overlapping structures. This overlap is a frequent cause of diagnostic errors and potential legal issues.
Case 3: Technical Issue
Poor image quality due to motion artifacts, incorrect positioning, or technical errors can hinder interpretation for both radiologists and AI. In some cases, AI may generate a weak “light blue” heatmap, indicating low confidence. The radiologists need to retrospect the image to verify whether the highlighted area identified by AI corresponds to a lesion or not.
Case 4: Confirming Pneumonia
Both the radiologist and the AI system agree on the presence of abnormal findings suggestive of pneumonia. This helps improve diagnostic confidence.
Case 5: The Growing Lesion
This case, initially missed by humans but detected by AI with increasing confidence as the lesion progressed, showcases the tremendous potential of AI in identifying subtle changes over time. Early detection, as demonstrated here, can be life-saving for patients with progressive conditions.
Pitfall of AI
- Accuracy
AI is not flawless. While it excels at identifying specific patterns, its accuracy can vary depending on the type of lesion or abnormality it’s trained to detect.
Case Example: The AI misclassifying a mediastinal mass as a pulmonary mass highlights the importance of understanding the limitations of specific AI models and not blindly relying on their outputs.
- Image Quality Matters
Poor image quality due to factors like motion artifacts or technical errors can significantly impact AI performance, leading to inaccurate diagnoses.
- Anatomical Variation
Variations in anatomy, like overlapping structures, can pose challenges for both radiologists and AI, requiring careful interpretation and collaboration.
- Difficult Cases Remain Difficult
Difficult cases that pose a challenge for radiologists will likely remain difficult for AI to interpret accurately.
