SRH: Imaging Technique Combined with Deep Learning Enables Quick Diagnosis of Brain Tumor during Surgery

Paper source: doi:10.1038/s41551-016-0027

The neurosurgeons and pathologists at Michigan Medicine recently developed a new approach, combining a powerful imaging technique with deep learning algorithm for automatic tumor diagnosis during brain surgery. It allows for a faster, safer and more accurate diagnosis in the operating room to distinguish between cancer and normal tissue. This paper was published in Nature Biomedical Engineering, 2017.

Motivation of the study
During surgery, neurosurgeons try to remove as much tumor as possible while retaining healthy brain tissue. However, they will wait for more than 30 minutes after removing a brain tumor for pathology results. During this period, the tissue is sent to a pathology lab, where it will be processed, sliced, stained and examined under the microscope. The neurosurgeons in the operating room needs to wait for the pathology results in order to make better decisions.

To shorten this process, the authors developed a new approach that allows for a quicker and more accurate diagnosis in the operating room, reducing it from 30 minutes to about 3 minutes. The new approach would shorten surgery time, save money, and reduce risks to the patient. The authors’ new method, called Stimulated Raman Histology (SRH), performs quick diagnosis within the operating room by combining images from SRH with a computer program that can automatically classify the images (Figure 1).

Figure 1. SRH system in the operating room.

SRH, the new imaging technique
The device behind SRH is the Stimulated Raman Scattering (SRS), a type of microscopy that does not need tissue processing, slicing or staining. However, the SRS microscopes are too big and expensive for an operating room. In addition, the lasers used in SRS are also unsuitable for use in an operating room. To this end, the authors created a portable, safe, and high-resolution system, called SRH, by switching to a fiber-laser microscope. The SRH can generate images similar to the traditional method (Figure 2). Thus, pathologists can easily interpret the resulting images from SRH and distinguish the tumor tissue from normal brain tissue.

Figure 2. Images from SRH and the traditional method (H&E) for brain tumor samples.

To validate the results from SRH, the images of specimen samples generated through both SRH and the traditional method were assigned to three neuropathologists for diagnosis. The neuropathologists were told the same information about each patient’s medical history and the location of the tumor. These neuropathologists made a near-perfect diagnosis regardless of whether they used SRH or traditional images. These results suggest that the SRH method is suitable for making accurate diagnosis (Figure 3).

Figure 3. Diagnosis results by neuropathologists using images from SRH versus traditional method.

Combining with deep neural network
Furthermore, the authors developed a computer program to perform SRH-image-based automatic diagnosis by using machine learning algorithm, called multilayer perceptron (MLP). The MLP is a feedforward artificial neural network model that is easy to iterate, powerful in distinguishing non-linear correlation, and efficient with current computational power. The MLP consists of three or more layers (eight layers herein) of nodes and is considered as a deep neural network (Figure 4).

Figure 4. The structure of deep neural network in MLP.

The authors generated 12879 images from 101 patients using SRH, and calculate 2919 image attributes for each image. These normalized image attributes were then put into the MLP for model training, iterating to minimize the difference between predicted and observed diagnoses through backpropagation algorithm. The MLP was programmed with two software libraries: Theano (http://deeplearning.net/software/theano/index.html) and Keras (http://keras.io).

The authors then used a leave-one-out approach to test the accuracy of the MLP model. The model can accurately predict the diagnostic subtypes for the samples, even for the samples with high histoarchitectural heterogeneity. In addition, the authors evaluated the accuracy of the model in a test set of samples diagnosed by the neuropathologists. Overall, the MLP model was able to predict brain tumor subtype with 90% accuracy (Figure 5).

Figure 5. The accuracy of prediction by MLP model. NL, non-lesional; LG, low-grade glioma; HGG, high-grade glioma; NG, non-glial tumour.

Future directions
The accuracy of the current model can be improved by feeding it more training data. In the future, the SRH system can assist with clinical decision-making during surgery, especially for medical centers without experienced pathologists on staff. In addition to brain tumor, the SRH system can be expected to be applied to other tumor types as well.

 

Reference
Orringer DA et al. (2017) Rapid intraoperative histology of unprocessed surgical specimens via fibre-laser-based stimulated Raman scattering microscopy. Nature Biomedical Engineering.

---

Analyst: Genome Hunter | Localized by Synced Global Team : Xiang Chen