The sophisticated HS Analysis Deep learning software is able to detect a plethora of sicknesses, one case is Chronic myeloid leukemia, (CML) is a type of aggressive cancer that begins in the bone marrow’s blood forming cells and attacks white blood cells and bone marrow, causing the immune system to weaken which spreads throughout the body. Specialists at HS Analysis are able to use advanced deep learning AI software to diagnose CML expertly along with its classes to aid clinics, institutions and health care facilities.
The HS Analysis‘ touch:
One key technology of automatic interpretation of tissue samples in the HS Analysis software is the latest artificial intelligence. Deep learning evaluates a statistical predictive model for CML analysis by using algorithms to „learn“ how to recognize CML disease and gain access to variables that predict CML survival. Some examples include quantification, diameter measurements, and the separation of healthy and unhealthy cells.
The HS Analysis software can help medical professionals diagnose disorders by examining high-resolution images of tissues and organs. And as the program generates data describing the patient’s current condition, it will assist doctors in selecting appropriate therapies, some examples of treatments will be discussed later in this blog.
The images below show the UI of HS Analysis KIT, in which the project name, tool bar and tabs, etc… can be seen.
Within the regions of interest, drawings with manual user annotations are made. After studying the annotations, the deep learning model applies them to several, larger regions of interest. The individual cell is annotated with respect to its type for example, the following images display that the Blood cells can be seen with Yellow, the healthy cells are distinguished by red whereas the un-healthy ones are Green. Their classes will be explained later on.
For a better and more interactive viewing, you can move the slider from left to right as well as zoom in and out to see the true supreme cell detection quality that our HS Analysis KIT can do on a small portion of a much larger whole slide picture of CML that was able to differentiate types of cells and detected the exact diameter of each cell. Below you can see the before and after CML healthy\un-healthy cell annotation using the automated deep learning instance segmentation HS Analysis KIT model.
Furthermore, in terms of classes, we can see the many different classifications of the un-healthy cells that are detected by the HS Analysis software, just like the other annotations, these classes must first be annotated as well before they are trained with the deep learning model, there are 40+ CML classes that the HS Analysis KIT can learn from and it differentiates by its names and colors of the annotated structure, for example they range from the young\old Megakaryozyt to the very common and thin spiders web like NET.
The main treatment for chronic myeloid leukemia is a tyrosine kinase inhibitor (TKI). Research facilities and pharmaceutical companies can test whether TKIs promote or decrease the formation of NETs, which are obtained from CML patients.
The quantity of mature\immature Megakaryozyt known as blasts, can determine which of the 3 phases the CML patient is currently in.
In terms of lab results, after the model is trained and ran the entire project, the HS Analysis software provides us with the information we need to understand the severity and condition in a sheet format, such as the name of the files, number of objects (structures) such as healthy and un-healthy cells, the classes of the un-healthy cells, dimension of the analyzed areas such as square meter and diameter of cells sizes and their average, etc…
Or we can specify the deep learning analyzation of a single structure like Healthy cells or one\multiple of its sub-structures such as Neutrophil Extracellular Trap (NET)s
By utilizing the proprietary deep learning HS Analysis software and its lab result, Doctors can have an ease in mind knowing that these accurate results can assist in making decisions for patients that suffer from Chronic myeloid leukemia or a research center that are seeking for a better insight into Chronic myeloid leukemia.
The report bone marrow morphology CMLpead will be perfectly integrated as a digital report into the Software HSA KIT. The numbers from the quantification with HSA KIT will be written into the report and are the basis for the evidence based medicine. Based on this report the doctors get an AI assistant and suggestions to the possible development and status of patient.
CML patients can rest easy knowing that the collaboration of the physicians and medical researchers at Universitätsklinikum Erlangen and the specialists in image analysis and artificial intelligence In digital pathology at HSA will make the route to discovering better medicine much more advanced and not to mention reducing the rate of mortality in chronic myeloid leukemia.
Most advanced instance segmentation for CML worldwide:
The software HSA KIT includes HyperCMLNet, the most advanced instance segmentation for CML worldwide. HyperCMLNet as deep learning technique was developed at HS Analysis GmbH and based on Mask R-CNN and Vision Transformer (ViT).
In the software HSA KIT data scientists not only create and manage high quality ground truth data or train and manage corresponding deep learning models, they also are able to compare the models with metrics as well as overlays.
You can see below the metrics of the first HyperCMLNet (type 1) based on Mask R-CNN as well as the second HyperCMLNet (type 2) based on Vision Transformer (ViT). The HyperCMLNet in the first step separate erythrocyte from leukocytes and in the second step separate all the classes from leukocytes. The doctors are able to detect fully automatically all the classes of CML e.g. NET, Pseudo-Gaucher-Zelle and Megakaryozyt, etc. as well as erythrocyte. The classes were trained on 137,834 GTD. These models were evaluated using both metrics mAP and loss in the software HSA KIT.
In the case of Loss, which is a function that requires to be decreased in order to have an ideal deep learning model. The loss value would produce a huge number if its predictions diverge too much from the actual results. The graph shows that the HyperCMLNet which is based on Mask R-CNN (Type 1) has a lower Loss when compared to the(Type 2) ViT based HyperCMLNet. We can see that the HyperCMLNet (Type 1) based on Mask-R-CNN has the loss metrics values of 3% and 11% when compared to the higher loss metrics values of HyperCMLNet (Type 2) based on ViT which are 4% and 13%. So in terms of Loss, the HyperCMLNet (type 1) has the best Loss evaluations out of both types.
In the case of mAP, the accuracy of the instance segmentation detection is measured using mean average precision (mAP), another performance evaluation metric. Thus, the outcomes of instance segmentation will be more accurate the higher the mAP value. The graph also shows that this time the HyperCMLNet (Type 2) based on ViT has superior mAP values being at 94% and 96%, when compared to the lower mAP values produced by the HyperCMLNet (Type 1) based on Mask R-CNN which are 86% and 89%. So in terms of mAP, the HyperCMLNet (type 2) has the best mAP evaluations out of both types.
Visual interpretation of the results:
After the AI training is finished for any model, it always ends in giving the user mathematical results such as graphs or numerical results. Although they are important, another determination of the quality of the AI model is done by checking the detection manually. Which leads us into this section of visual interpretation of the results, and this section illustrates the visualization the instance segmentation of both of the (type 1 & 2) AI models of the CML images, the following figure shows multiple images are showing the original and the (predictions) determined by the networks.
The images of cell detection acquired from model trainings in the accompanying figure ((A) Erythrocyte and Leukocytes, (B) NET, (C) Megakaryozyt, and (D) Pseudo) demonstrate that both (types 1 & 2) of HyperCMLNet have very similar detection. First off, as the name suggests, the original picture column displays the untrained, raw, and cell-detection-free portion of the CML image. The HyperCMLNet trained model has been applied to some of the Base ROIs shown in the second column (Type 1). These Base ROIs display the detection of erythrocytes and leukocytes as well as the substructures of leukocytes.
Thirdly, the (Type 2) HyperCMLNet trained model is displayed in the third and final column and was applied to the identical Base ROIs as the (Type 1) model. When type 1 and type 2 are compared in terms of erythrocytes and leukocytes, both models exhibit excellent detection of erythrocytes and leukocytes, possessing a high level of detail and detection boundaries.
The HyperCMLNet architecture’s brilliance was able to distinguish the error prone when detection of the separation of some mashed-up erythrocyte cells as well as leukocytes, even though their error occurrence is less likely because they have a significantly smaller area especially in comparison to erythrocyte. This project was both large and complex. In order to distinguish between closely spaced erythrocyte cells, the HyperCMLNet was able to deliver its findings in a clear and understandable manner.
Again, the quality is remarkable in both (type 1 & 2) of the HyperCMLNet designs when comparing the (type 1 & 2) models in terms of NET identification. It has the ability to ignore non-NET cell detection and focus only on NET cell detection.
Types 1 and 2 are contrasted with Megakaryocyte and Pseudo. Both (type 1 & 2) seem to detect Megakaryocyte and Pseudo with high precision and accuracy in all files and delivering outstanding detection. This is because both (type 1 & 2) employed high volumes of data for the HyperCMLNet (type 1 & 2) training. However, in all most all cases, the HyperCMLNet (type 2) showed a bit more attention to detail when detecting the CML structures when compared to the HyperCMLNet (type 2). The figure below shows the visualization of the training results of both types of HyperCMLNet.
HSA KIT’s Advanced Heatmap tool:
The enhanced heatmap tool detection that was utilized in HSA KIT will be employed in this section to compare the outcomes of the CML classes to one another. The advanced heatmap tool in the HSA KIT is used to detect the forecasting where the classes of the AI models are most likely to be located. The warmer (more red) the heatmap, the higher the probability of the detection; conversely, the colder (more blue), the less likely the detection prediction will occur.
In essence, the hotspot of AI-generated annotations is visualized using the sophisticated heatmap tool. It is designed for AI visualization projects like the one used in this study that produce findings with a lot of Detail displayed on a single slide.
The HyperCMLNet-ViT (Type 2) and HyperCMLNet-Mask-R-CNN (Type 1) architectures are used in the first column and the erythrocyte, leukocyte, and (B, C, and D) leukocyte classes, respectively, in the second column of the following figure.
We can’t really say which of the two architectures in (A) is more ideal because there isn’t much of a difference between them in terms of advanced heatmap detection of erythrocyte and leukocyte mother structures, but both architectures have great hotspot predictions.
When it came to (B, C, and D), the detection quality for both designs was comparable and yielded encouraging results, although they shared a lot of the same traits when it came to precise and accurate heatmap detection..
With the help of the advanced heatmap tool in the HSA KIT, the localization of the heatmap prediction can be further enhanced by adding more GTD. The advanced heatmap tool also offers the ability to set the intensity slider so that we can view the heatmap at various intervals and fully appreciate the revolutionary detection of the HyperCMLNet architecture. The figure below shows the visualization of HSA KIT’s advanced heatmap tool for the training results of both types of HyperCMLNet.
Digitalization of histological slides and analysis with HSA KIT
HSA imports and works with 3dhistech and mirax files created by the Hamamatsu slide scanner. They are a perfect match for the HSA KIT.
If you don’t have a slide scanner and want to get one later, you can manually digitalize the slides on your microscope and create manual WSI using our inexpensive and affordable software (HSA SCAN M).
In this video, you can see manual scanning that converts a slide to a file with HSA SCAN software so that it can be trained and then automatically detected by a trained AI model using HSA KIT software.
If you want to automatically scan the slides, you can upgrade your microscope to an automated microscope station (HSA SCAN A). This converts your microscope into an automatic scanner for low-cost, high-quality performance in a short period of time.
HSA only requires the dimensions of your microscope and any specifications you wish to include, and we will send a stand and motor for your device that fits your microscope.
This will help you in:
- Scanning faster
- More accurate scanning
- Saving time
- Low budget
In above video, you can see an example of automatic scanner (HSA SCAN A). Do not hesitate to contact us for further information or ordering.
For more information or ordering : email@example.com
Note: This website will be updated in future. More CML healthy cells sub-structures will be added to the modules.