Article tittle A deep learning method and device for bone marrow imaging cell detection

Author Jie Liu, Ruize Yuan, Yinhao Li, Lin Zhou, Zhiqiang Zhang, Jidong Yang, Li Xiao

Methodology In this article, They establish a new morphological diagnosis system for bone marrow cell
detection based on the deep learning object detection framework. The model is based on the
Faster Region-Convolutional Neural Network (R-CNN), a classical object detection model.
The system automatically detects bone marrow cells and determines their types.
Asspecimens have severe long-tail distribution, i.e.,the frequency of different types of cells
varies dramatically, They proposed a general score ranking loss to solve such a problem.
The general score ranking loss considers the ranking relationship between positive and
negative samples and optimizes the positive sample with a higher classification probability
value.
1.Data Collection and Preparation:-one marrow specimens from 70 leukemia patients
were used for digital scanning of whole slides to establish a database of bone marrow
morphological characteristics.
2.Data Augmentation:-mage horizontal flip, vertical flip, rotation, translation, and adding Gaussian
noise.
3.Model Development:-The model was developed based on the Faster R-CNN framework, a classical
object detection model.
4.Integration with Clinical Diagnosis Indicators:The deep learning model was combined
with expert diagnosis indexes of clinical bone marrow morphology.
5.Hardware Integration:The software developed was integrated into the microscope
system to create an augmented reality system for bone marrow cell detection.
6.Evaluation:The response speed of the diagnostic system was compared to that of trained
diagnostic experts through clinical tests.
 Cont...
Result  They developed their model based on the Faster R-CNN. In view of the decline of
recognition
accuracy caused by many types of bone marrow cell images and unbalanced sample
distribution, a generalized average precision loss (G-AP loss) was designed so that the model
could decrease negative samples effectively in training. Finally, combined with the diagnosis
indicators of clinical bone marrow morphology experts, the results were systematically verified
and tested, and a new type of hematological disease diagnosis model was established.

Discussion  They developed a new morphological diagnosis system for bone marrow cells using the
Faster R-CNN model. To handle diverse feature distributions, they enhanced data
augmentation by combining five methods: horizontal flip, vertical flip, rotation, translation,
and Gaussian noise addition. Addressing the issue of highly imbalanced data, they designed a
generalized rank loss for better optimization of positive samples. their model has been
successfully integrated into the software and hardware of an augmented reality microscope,
showcasing significant potential for clinical use.

Gap The limitations of article is bias towards a particular population and difficulty learning rare
samples.
i.e Although experiments demonstrate the significant performance improvement of their
proposed methods over most of the metrics, their model is still biased towards a single center
and has to eliminate rare samples before training.

Criticism Its limited discussion on the generalizability of the deep learning model beyond the specific
dataset from The Second Affiliated Hospital of Zhejiang University. Although the study reports
high accuracy, it lacks validation on external datasets and real-world clinical scenarios, raising
concerns about robustness and applicability. The article also does not address challenges in
implementing the automated WBC counting system in practical clinical environments, such as
integration with existing workflows, scalability, or regulatory issues. Further research in varied
clinical settings is needed to assess real-world performance and feasibility
 Cont...
Article tittle Deep learning for bone marrow cell detection and classification on whole-slide images

Author Ching-Wei Wanga,, Sheng-Chuan Huang, Yu-Ching Lee , Yu-Jie Shena,Shwu-Ing Meng, Jeff L. Gaol

Methodology The methodology of the article by Ching-Wei Wang et al. involves the development and
implementation of an efficient and fully automatic hierarchical deep learning framework for bone
marrow nucleated differential count (NDC) analysis on whole-slide images (WSIs) using a 40x
objective magnification.
They develop an efficient and fully automatic hierarchical patch-based deep learning framework for
BM NDC WSI analysis in seconds. The proposed hierarchical patch-based deep learning framework
with two Cascade R-CNN deep learning models usesa coarse-to-fine strategy to rapidly locate BM
particles and cellulartrails and identify BM cells in WSIs by accessing two magnification levels
The key steps and components of the methodology include:
1.Deep Learning Model for Rapid Localization: The framework includes a deep learning model
designed for the rapid localization of bone marrow particles and cellular trails on WSIs. This initial
step generates regions of interest (ROI) for further analysis.
2.Patch-Based Deep Learning Model for Cell Identification: A patch-based deep learning model is
utilized for the identification of 16 different cell types within the bone marrow samples. This includes
cell types such as megakaryocytes, mitotic cells, and various stages of erythroblasts.
3.Fast Stitching Model for Integration: A fast stitching model is employed to integrate the patch-
based results and produce final outputs for analysis. This step ensures a comprehensive and cohesive
analysis of the bone marrow samples.
4.Evaluation: The proposed methodology is evaluated using datasets containing annotated cells. The
evaluation includes measures of recall, accuracy, and computational time to assess the performance of
the deep learning framework in bone marrow NDC analysis.
5.Comparison with Benchmark Methods: The methodology involves comparing the performance of
the proposed deep learning framework with existing small-image-based benchmark methods to
demonstrate its superiority in terms of recall, accuracy, and efficiency.
 cont...
Result

Discussion In this study, They present a hierarchical deep learning framework for analyzing BM NDC WSI in
seconds. The fully automatic and efficient system can reliably differentiate BM cells, a critical task in
diagnosing and managing hematologic disorders. The framework detects BM cellular regions in WSIs
without examiner annotations and identifies BM particles and cellular trails, performing cell detection
and classification. It is applicable to pathology and cytology digital slides from various specimens.
Notably, their study explores detailed stages of erythroblasts and other important cell types in BM for the
first time.
Gap The gap is the limited size of datasets used for training and testing the deep learning model for bone
marrow cell detection and classification. With datasets collected from a small number of patients, a
larger and more diverse dataset could improve the model's robustness and generalizability by capturing
the variability in bone marrow samples across conditions. External validation datasets from various
sources could validate the model's performance across populations and settings. Addressing these gaps
through dataset expansion and external validation could enhance the study's findings and the clinical
utility of the deep learning model for bone marrow analysis on whole-slide images.
Criticism An article for lacking detailed discussion on the interpretability and explainability of a deep learning
model used for bone marrow cell detection and classification. It highlights the importance of
interpretability in medical applications to build trust among clinicians. The critique suggests that
addressing how the model's predictions are interpreted, aligning its decisions with medical knowledge,
and discussing biases or limitations could enhance transparency and trustworthiness in clinical settings,
ultimately improving the model's impact on bone marrow analysis in healthcare.