Medical Image Analysis Using W&D (Wide And Deep Network Model)

3 main points
✔️ Improve the image analysis and diagnosis technology of prostate cancer using W&D (Wide and Deep Network Model)
✔️ Achieve image classification with higher accuracy than existing machine learning models
✔️ Prove the high versatility of W&D

Wide & Deep neural network model for patch aggregation in CNN-based prostate cancer detection systems
written by Lourdes Duran-Lopez, Juan P. Dominguez-Morales, Daniel Gutierrez-Galan, Antonio Rios-Navarro, Angel Jimenez-Fernandez, Saturnino Vicente-Diaz, Alejandro Linares-Barranco
(Submitted on 20 May 2021)
Comments: Accepted by arXiv
Subjects: Machine Learning (cs.LG)

code：   

first of all

Over the past few years, AI technology for image recognition has made significant progress. One of the reasons for this is the expanded use of Deep Learning (DL).

In this article, we compare the diagnostic accuracy of various models of medical image analysis through the medical image diagnosis of prostate cancer (PCa), and among them, we focus on W&D (Wide & Deep Network Model), which is a network that was once a hot topic in the recommendation, and utilize it for medical image analysis. I will explain the paper "Wide & Deep neural network model for patch aggregation in CNN-based prostate cancer detection systems".

Prostate cancer (PCa) is the second most frequently diagnosed cancer in men and the fifth most common cause of cancer death, accounting for more than 1.41 million deaths in 2020 and approximately 375,000 deaths worldwide.

1. In general, the first step in diagnosing PCa is to perform a digital rectal examination (DRE), which is the primary test for initial clinical evaluation of the prostate.
2. If an abnormal DRE result is found, prostate-specific antigen (PSA) analysis is performed as a screening method for tumor investigation.
3. If the PSA is positive, a transrectal ultrasound-guided biopsy may be considered.
4. Prostate samples are acquired, processed, and scanned in the laboratory to produce gigapixel resolution images called whole slide images (WSI). These images are then analyzed by a pathologist to ultimately diagnose whether or not it is cancer.

In this medical image analysis, the WSI of prostate tissue is learned to analyze and detect whether it is normal or malignant with high accuracy.

These algorithms may play an important role as a second opinion and as a screening method to assist physicians in specific image analysis tasks. In particular, convolutional neural networks (CNNs) and various other DL mechanisms have recently been learned to analyze and detect malignant tissues in WSI and are widely used in digital histopathology.

Since the results obtained using CNNs are reported at the patch level, different techniques have been proposed to combine them to generate slide-level classification results. This technique is known as patch aggregation.

Among the various studies proposed, various patch aggregation techniques based on Recurrent Neural Networks (RNN), Random Forest (RF), other Machine Learning (ML), and statistical alternatives have been implemented, which can help the pathologist to work out an accurate leading to screening method.

Data sets and preprocessing

A set of hematoxylin and eosin (H&E) stained slides (158 normal WSI and 174 malignant WSI) provided by the Pathological Anatomy Unit of Virgende Valme Hospital (Seville, Spain) is used. These images are subjected to patch-level processing of WSIs to prediction by the following procedure

1. In a step called Read, patches are extracted from the input WSI and the corresponding patches in the background are discarded.
2. In the Scoring step, a score is assigned to each patch in order to discard patches that correspond to unwanted regions. This score is differentiated by considering three factors: the amount of tissue in the patch, the percentage of pixels that fall within the H&E hue range, and the variance of the saturation and brightness channels. Patches discarded in this step are shown in red, while patches that pass the scoring filter are highlighted in green.
3. In the third step, called Stain normalization, color normalization is applied to the patches based on Reinhardt's stain normalization method to reduce the color variation between samples.
4. In the Predict step, each patch is used as input to a CNN called PROMETEO, which classifies them as either malignant or normal tissue.

PROMETEO is a CNN-based computer-aided diagnosis system for WSI prostate cancer detection.

Malignant (left) and normal (right) samples are also distinguished in the following figure.

This histogram is the mean probability histogram of normalized patch frequencies across all WSIs. The red dashed line represents the least squares regression line.

From the histogram, in the cancer WSI, the patches classified as malignant tend to be concentrated in the tumor area. Therefore, we hypothesize that the dispersion coefficient of patches classified as malignant could be another input for classifying normal and malignant WSI at the sliding level. This coefficient was obtained by calculating the number of malignant connected components (MCCs).

See Algorithm 1 for details on how to calculate the number of connected components based on the center coordinates of the malignant patch.
This value corresponds to the Euclidean distance (radius) from the patch and ranges from 1 to 5 patch distances, considering that the distance between two patches is 100 pixels (the size of the patch is 100 × 100 pixels ). The number of connected components is normalized with respect to the total number of malignant patches in each WSI.

Algorithm 1: Algorithm for finding the MCC

W&D

As the name suggests, the W&D model has both a wide component (width) and a deep component (depth). The wide component effectively stores the sparse interactions between features. It can be defined as learning how the output responds to a combination of sparse input values.

The deep component, on the other hand, corresponds to a forward propagating neural network and represents generalization, i.e., the ability to handle data that has never been seen before. Therefore, it combines the advantages of both memory (wide) and generalization (deep) in a single model, combining the advantages of both memory (wide) and generalization (deep)

Learning and Verification

The W&D network model used in this verification is as follows

Each hidden layer is composed of 300 neurons.

The input is

• Malignant Tissue Ratio (MTR) of WSI
• Slope and Y-intercept of the least squares regression line (LSRL) of the histogram
• (MCC) is omitted because it is described in detail above.
• 10-division Malignancy Probability Histogram (MPH) from 50% to 100% with 5%ticks

These inputs are used to classify the WSI as either malignant (M) or normal (N).

From here on, the entire processing steps of the entire PCa screening task

First, we process the WSI at the patch level, following the same steps shown in the preprocessing.
Then, using the output classification of each filtered patch from the original WSI, we perform a slide-level prediction using the W&D model shown in the figure above.

To measure the generalization performance of the model, K-fold stratified cross-validation (K-split cross-validation using stratified extraction method) is performed.To validate the network, we use different evaluation metrics: accuracy, sensitivity, precision, F1 score, and AUC of the F1 score-ROC (ReceiverOperating Characteristic) curve.

Results and Discussion

The results of K-fold stratified cross-validation and the average values of all evaluation indices are summarized as validation results of W&D's network model.

The mean results were ACCURACY 94.24%, SENSITIVITY 98.87%, PRECISION 90.23%, F1 score 94.33% and AUC was 0.94. It achieved very high scores in all the metrics considered in this classification task, especially in terms of Sensitive. Sensitive is defined in this field as the ability of a system to identify PCa and is of paramount importance in reporting and evaluating the performance of screening tests. This means that the W&D model is a reliable patch aggregation method, with almost no errors in predicting malignant samples.

Here are the results compared to other well-known machine learning models

As can be seen from the table, the best results are obtained by the proposed W&D model for Accuracy, Sensitivity, F1score, and AUC, with the exception of Presicion, where SVM achieves the highest value.

As mentioned before, W&D is the best in terms of image analysis in this case since Sensibility is the most relevant metric to measure the performance of a classifier when screening. SVM achieves a Presicion of about 99%, which is considered to be very suitable for other binary or multiclass classification tasks.

summary

We found that the W&D model outperforms other machine learning models in multiple aspects: accuracy, sensitivity, F1score, and AUC.

This W&D model is not only a recommender function but also a useful algorithm in image analysis, which may be very useful for pathologists to analyze histopathological images as a screening method and identify normal PCa and malignant PCa slides.

However, it is important to note that although this is still a patchy study, I am not convinced that we have a sufficient amount of data yet. Therefore, as more data is added and the number of data increases, it will be necessary to compare with deeper models in the future. I think W&D is good enough to have shown one possibility this time.

• 
• 
• 

Categories related to this article

• Machine Learning
• Deep Learning
• Image Recognition And Analysis
• Segmentation
• Object Detection
• AI Vulnerabilities
• Transfer Learning