AI-Driven Deep Learning Model Shows Promise in Standardizing MDS Diagnosis

According to Palak Dave, distinguishing myelodysplastic syndromes (MDS) from pre-MDS conditions like idiopathic cytopenia of undetermined significance (ICUS) and clonal cytopenia of undetermined significance (CCUS) is a main challenge in the space. While MDS and pre-MDS conditions share similar abnormalities, pre-MDS ones have a milder presentation that makes diagnosis subjective and prone to variability, leading to inconsistent categorization.

Dave, postdoctoral fellow at Moffitt Cancer Center and Research Institute, addressed these challenges in a presentation of an innovative study at the 2024 American Society of Hematology (ASH) Annual Meeting and Exposition. The study explores how artificial intelligence (AI) can help standardize and accelerate MDS diagnosis.

Using a deep learning model that analyzes bone marrow aspirate smear images, this approach has the potential to reduce subjectivity and improve diagnostic accuracy. Early results show promise, with AI achieving about 70% accuracy in internal validation.

Further research is needed to validate this AI method with larger datasets and incorporate other clinical factors, such as bone marrow biopsy results, to improve its effectiveness.

“More than accuracy, it is about standardizing and making things faster. I do not think human experts will be redundant anytime soon, but when they work together, AI and human experts, they can create a greater sum than their individual parts,” explained Dave in an interview with Targeted Oncology^TM.

In the interview, Dave further discussed the challenges in diagnosing MDS vs pre-MDS conditions and findings from this research presented at ASH 2024.

Targeted Oncology: What challenges exist in diagnosing MDS vs pre-MDS conditions?

Dave: The main difference between MDS and pre-MDS conditions lies in the threshold of abnormalities. In pre-MDS conditions, such as ICUS and CCUS, they do have abnormalities similar to MDS, but not as high as MDS. When they analyze these things manually, there are challenges, [particularly due to] subjectivity. This subjectivity can result in variability, potentially causing a patient to be categorized differently. To address this, we are working to standardize the process to ensure more consistent and reliable classification.

Red blood cells flowing realistically, 3D rendering: © Kodjovi - stock.adobe.com

Can you provide an overview of the abstract that is going to be presented at ASH on this topic?

It is a deep learning approach designed to automatically distinguish MDS from pre-MDS conditions using only bone marrow aspirate smear images. No additional clinical data, genetic testing, or other inputs are involved—just image data. The goal is to explore how accurately this method can perform using only pathology images.

What role could AI play in improving the accuracy of MDS diagnosis?

I would say the focus is not just on accuracy but also on standardization and speed. Experienced hematopathologists can achieve high diagnostic accuracy, but the process can be time-consuming. Additionally, there is inherent subjectivity among experts, which can lead to variability in diagnoses. By standardizing these processes, we can make them faster and more consistent, ultimately benefiting patients.

How does this AI pipeline standardized bone marrow smear analysis?

So I can go into a little more depth when I say there exists subjectivity. Why does that exist? Basically, in these kinds of diseases, the hematopathologist looks at individual cells in the bone marrow aspirate smear and makes a decision about what category the cell belongs to, out of many categories and possible abnormalities. So in this visual review, there can be subjectivity. I may think it is Category A, and another person may think it is Category B, because the differences are very minute. So when a machine does it, it is kind of standardized.

What are the key features of the AI models used in the study?

Right now, we are using a cascade of models. There are 3 models: First is classification model, second is the detection model, and third is, again, a classification model, but it is a more recent, advanced model.

How effective was the AI in distinguishing MDS from non MDS cases? What are the potential benefits of AI for community oncology practices?

In terms of accuracy on our internal validation set of about 129 patients, it is about 70% accurate at this time. On an external test set with 31 patients, it is accurate for about 65%.

More than accuracy, it is about standardizing and making things faster. I do not think human experts will be redundant anytime soon, but when they work together, AI and human experts, they can create a greater sum than their individual parts.

What additional research is needed to validate this AI approach?

This is a very early study in this field. I think this approach itself needs to be tested on a larger external dataset so that we can confirm it's generalizable and not just fitting the data used for training. That's the first step. After that, we plan to combine other aspects of these patients, like bone marrow biopsy, or potentially integrate some clinical features going forward.

For a community oncology audience, what are the key takeaways of this abstract?

I would say it is not the first, but one of the very early studies that examines whole slide images and provides automatic diagnosis of a patient. There have been studies that analyze individual cell types or focus on small parts of the overall pipeline, but there is no existing pipeline that goes to patient-level diagnosis. So, the key takeaway is that this is the kind of study that addresses that gap.

REFERENCE:

1. Dave P, Rai J, Zhang L, et al. A deep-learning pipeline for diagnosis of myelodysplastic syndromes/neoplasms using bone marrow smears. Presented at the 2024 American Society of Hematology (ASH) Annual Meeting and Exposition; December 7-10; San Diego, CA; abstract 4991.