Before artificial intelligence was integrated in the medical workflow, doctors used to spend a lot of time analyzing medical images to reach a certain diagnosis because it was done manually and was also not that accurate.
We as experts in HS Analysis strive to provide the best healthcare and treatment of various diseases, including ophthalmic diseases, by utilizing Artificial Intelligence and especially Deep Learning in hospitals, laboratories and research centers. Identifying the disease is a challenge for the ophthalmologist, but this problem can be solved using AI as mentioned earlier. This is done by a skilled and experienced staff, using advanced hardware and software to provide rapid and accurate diagnosis of ocular diseases, including some rare diseases like neurotrophic keratopathy.
How can HSA KIT SOFTWARE help ophthalmologists?
NK can lead to severe complications and even blindness. Identifying the agent in the early stages and preventing complications is a challenge for the ophthalmologist. Here in HS Analysis, we manually annotate images that we receive from research centers and hospitals using our specific annotation tools. by annotating, we obtain helpful data that can be utilized by the AI in our sophisticated software to learn how to automatically detect and annotate specific characteristics in the received images in the future. The final results yielded by our software can tremendously help the health care section to better diagnose and treat ocular diseases. Corneal changes that our software can detect include:
- A significant reduction of nerve density in the sub-basal nervous plexus,
- a lower epithelial and endothelial cell density,
- an increase in dendritic immune cells.
Invivo confocal microscopy with HSA KIT
The Heidelberg Engineering HRT3 is a compact ophthalmic device that uses confocal scanning laser microscopy to provide high-resolution images of the cornea, the conjunctiva or the limbus at the cellular level. You can scan and analyze the images captured by this device using the HSA KIT to generate a highlighted and annotated output image of branches and the main fibers of nerves in high quality and short time. IVCM images are some of the ophthalmological images that we have applied AI to them to determine the health, sensitivity and density of the corneal nerves in a quick and accurate fashion. NK, which is one of the diseases that affects the corneal nerve which can cause severe complications and even blindness, can be diagnosed with applying AI on an IVCM images of a NK patient. The earlier the disease is diagnosed, the better the outcome and the complication
Creation ground truth data
Here in HS Analysis, experimentation is conducted on some special deep learning architecture, which is continuously improved by the staff. The corneal nerve consists of 5 layers, with one of them being the subbasal layer that contains two structures: the main nerve fiber and branch. Both of these structures form a major structure known as nerve fiber. To train a deep learning model, ground truth data is crucial. The number of GTD files used consisted of 210, with about 30 files used for testing the models created. Making GTD involved several steps. Firstly, BMP images with a size of 384×484 were uploaded to the HSA KIT software. Then, the whole image was selected as the region of interest (ROI) because, for NK, images were treated as a whole image. Next, nerve cells were classified into two classes, main fibers, and branches, and annotated using the HSA KIT tools.
Interpretation of trained model results
Initially, The topic of nerve fibers will be covered first . 3 deep learning models were in house created using 3 different architectures which are (HyperNKNet50, HyperNKNeXt50, HyperNKNet-B3). As shown below, HyperNKNet-B3 was the best model for the nerve fiber because the loss value is lower in this particular architecture than the other models, and the accuracy was better than the other architectures.
The same experiment was repeated, but this time, More GTD were used to determine whether the model would benefit from it and whether better results would be obtained. According to the results, the accuracy of all models has increased, but only HyperNKNet50 and HyperNKNeXt50’s loss has declined, and this decline is harmful to the performance of the model. Increasing the GTD didn’t cause any significant change in the IoU of all three models. Generally, HyperNKNet-B3 performed the best in this testing as well.
Now, the classifications of nerve fiber will be discussed, which are main fiber and branch. The same architectures as the previous experiment were used to train the GTD. From the results of the three models, it can be determined that HyperNKNet-B3 was the worst model for classifying nerve fibers due to high loss values and lower prediction accuracy and IoU than other architectures. When comparing the two remaining models, HyperNKNet50 was determined as the superior model to HyperNKNeXt50 because it had lower loss values and better accuracy
The same experiment was performed again on the nerve fiber classification, but this time more GTD were utilized to see if the model would benefit from it and whether we would get better results. Interpreting the findings of our three models showed that HyperNKNet-B3 was the weakest model for the determination of nerve fibers due to its low (accuracy and IoU) values and large loss values compared to other architectures. HyperNKNet50 was the best model when compared to the other two architectures since it was more accurate and had lower loss values than HyperNKNeXt50. When more GTD were used, the HyperNKNet50 became more accurate and superior to what it was before, but other architecture, particularly the HyperNKNet-B3, didn’t get any significant benefit from a larger number of GTD.
Interpretation of xAI
Initially, The topic of nerve fibers will be covered first . 3 deep learning models were in house created using 3 different architectures which are (HyperNKNet50, HyperNKNeXt50, HyperNKNet-B3). As shown below, HyperNKNet-B3 was the best model for the nerve fiber because the loss value is lower in this particular architecture than the other models, and the accuracy was better than the other architectures.
Another option available in the xAI to determine sensitive areas in an image is heatmap. Two images were prepared of nerves affected with NK and grad-cam was applied on these images. The images below show regions with high density and sensitivity of nerve fibers as red, and regions with lower density and sensitivity of nerve fibers as colors other than red according to the degree.
In general, the Grad-CAM, which was developed using the HSA KIT software, has the advantage of clearly displaying the nerve fibers and determining the degree density and sensitivity also you can easily find the number of nerve or find that nerve was not appear.
The figure above displays the sensitivity and density of corneal nerve which was determined by the application of the activation map.
Summary and Outlook
In HSA KIT, pixel-based segmentation and classification of annotated corneal nerve was performed using Deep Learning Also, Grad-CAM was used to visualize the CNN to determine the density and sensitivity of nerves in HSA KIT with module Neuro keratopathy image and how the network makes these decisions. The purpose of this work is to compare the different Deep learning architectures and to determine which architecture is the best to correctly diagnosis this disease and also determine the effect of the amount of GTD on training the HSA KIT with module Neuro keratopathy. These findings strongly correlate with NK. With the help of our software, we will be detecting these changes, and provide useful data for the ophthalmologist to make an accurate and rapid diagnosis so the treatment can be started. It can also be used for monitoring the response to the treatment.
