This work by Islam et al presents a label-free AI/ML based framework to classify a subset of lymphocyte images (T4, T8 and B) collected using bright-field microscopy. The dataset used in the study (LymphoMNIST) is new and contains 80,000 images. They applied a teacher-student model with knowledge distillation and implemented this model directly on FPGA, achieving a low latency of sub-25 μs.

The work is novel and presents an exciting advancement in FPGA-accelerated machine learning for label-free cytometry. However, the biological implications are modest, and several aspects of the benchmarking strategy require strengthening before this can be considered for publication.

Here are my detailed comments.

1. The three chosen lymphocytes are well-characterized and does not offer much new insights. The dataset, even though large, has only these three classes and does not offer much variability. Some demonstration of how the correct and rapid identification of these cells could have downstream impacts (such as a real-time sorting demonstration) would be valuable and can highlight the real impact of this work.
2. It is also mentioned that the dataset contained both young and aged populations. Some analysis of whether there is an impact of age should also be presented.
3. How would you assess whether the model is memorizing the data, or it is learning something new? Can you check similarity across the training and test sets?
4. The model is only compared against ResNet based approaches. Testing against more recent models (such as vision transformers and EfficientNet) could provide further validations.
5. Some discussion about the biological implications for the sub-25 μs latency would be very useful.
6. The statistical analysis shown is very limited and a detailed analysis showing whether a model is significantly better than another is needed.
7. A flowchart detailing all the steps of the data pre-processing pipelines and a breakdown of the individual class distributions in training, validation and test sets need to be shown.
8. Why does the confusion matrix in figures 1 and 2 add up to 5100 when validation and test sets had 8000 images? This discrepancy should be clarified.
9. Inconsistent figure references such as “Figure ??(a)” should be corrected.