Keywords
Deep learning - Artificial neural network - Convolutional neural network - Nonlinear
modeling - Colon cancer
Artificial intelligence (AI) is a computer technology for mathematical modeling that
uses nonlinear statistical analysis as compared to linear relationship evaluated during
conventional prediction systems such as logistic regression analysis to identify the
relationship between input and outcome variables using pattern recognition techniques
([Fig. 1A] and [B]).[1] Several clinical outcome predictions are done using multilayer perceptron (MLP)
networks ([Fig. 1C]).[1]
[2] Since perception at different levels within the network (the hidden layers) leads
to correct prediction or classification, this is called the MLP network ([Fig. 1C]). This is quite similar to the function of human brain ([Fig. 1D]). In contrast to MLP, the lesion identification and its characterization from various
images is done by convolutional neural network (CNN) technology, which uses multiple
filters to classify the data.
Fig. 1. Schematic diagrams showing the principles of linear (A) and nonlinear relationship (B) between the data, function of multilayer perceptron network (C), which is akin to that of human brain (D).
Artificial neural network (ANN) is a form of AI that utilizes mathematical models
having structure and functions somewhat similar to human brain ([Fig. 1C] and [D]).[1] The most commonly used network in medical science, the MLP network, works in the
following manner.[1] First, the network gets trained using the data from patients, whose outcome are
known to the network. The network attempts to make prediction from the data that are
fed to it as input variables, which is called the feed-forward network.[1] If the prediction becomes incorrect, by back-propagation, the network adjusts the
weight of the interconnections to provide more correct prediction ([Fig. 1C]).[1] Advantages of AI as compared to conventional statistical methodology such as logistic
regression are presented in [Table 1].[2]
Table 1
Advantages of artificial intelligence (AI) as compared to conventional statistical
methodology
|
Parameters
|
Logistic regression based models
|
Artificial intelligence-based models
|
|
Modeling principle
|
Linear model
|
Nonlinear models
|
|
Number of parameters as input variables
|
Limited number, typically those significant on univariate analysis
|
Any number
|
|
Dynamicity
|
Nondynamic. Once developed, it does not improve further
|
Dynamic (it never stops learning)
|
|
Floor and ceiling effect
|
It is an usual limitation
|
No such limitation
|
|
Predictive accuracy
|
Less
|
More
|
|
Robustness and external validity
|
Usually less
|
Usually more
|
However, for identification of an image and recognition of its pattern using AI is
undertaken by CNN rather than MLP.[2]
[Table 2] lists the differences between CNN and MLP.[2] In this issue of the journal, Afzalpurkar et al reviewed the impact of AI in colorectal
polyp detection and characterization.[3] This is an upcoming area of gastroenterology, which has already reached the doorstep
of practicing clinicians and in the near future, it may bring a paradigm shift in
clinical practice.[4]
[5]
Table 2
Differences between multilayer perceptron (MLP) and convolutional neural network (CNN)
|
Feature
|
MLP
|
CNN
|
|
Input data
|
Suitable for structured data, such as tabular data, where the relationships between
features are not spatially organized
|
Designed for grid-like data, primarily images, where spatial relationships between
pixels are crucial
|
|
Architecture
|
This feed-forward neural network consists of an input layer, one or more hidden layers,
and an output layer. Each neuron in a layer is connected to every neuron in the subsequent
layer, and there are no connections within the same layer
|
It includes convolutional layers, pooling layers, and fully connected layers Convolutional
layers use filters to scan across input data, capturing local patterns, and pooling
layers reduce the spatial dimensions
|
|
Parameter sharing
|
Each neuron in a layer is independent of the others, and there is no parameter sharing
between them
|
Convolutional layers use shared weights (filters) across different spatial locations.
This allows the network to learn spatial hierarchies of features
|
|
Translation invariance
|
Lacks translation invariance, meaning it may not perform well when presented with
the same pattern at different locations in the input
|
CNNs exploit local connectivity and parameter sharing, leading to translation-invariant
features. This is especially useful for tasks like image recognition
|
|
Use
|
Commonly used for tasks like tabular data regression or classification where the input
features are not spatially correlated
|
Well-suited for image classification, object detection, and other tasks involving
grid-like data
|
|
Training data
|
Often requires a large amount of labeled data to generalize well
|
Can leverage pretrained models on large image datasets and fine-tune for specific
tasks with smaller data sets
|
|
Strengths
|
Versatile, can be used for a wide variety of tasks
|
Very effective for tasks that involve spatial data
|
|
Weaknesses
|
Not well-suited for tasks that involve spatial data
|
Not as versatile as MLPs
|
A busy endoscopist performing multiple colonoscopies throughout the day is expected
to be tired toward the later part of the day.[6]
[7] In the aviation sector, international law does not permit the pilots to be put on
duty beyond a particular number of hours each day; in contrast, doctors have to continue
to discharge their duties continuously over unlimited number of hours. Studies have
demonstrated that colonoscopies done in the afternoon were more often incomplete and
missed findings than those done in the morning hours.[6]
[7] Can machine-driven rather than human-driven system bring a solution to it in the
face of lack of adequate doctor–patient ratio in many countries not allowing limiting
the number of hours of doctor's duty immediately? The initial AI-assisted colonoscopy,
GI Genius, was developed by Medtronics with annotated colonoscopy videos obtained
from a colonoscopist and several AI experts working with gastrointestinal (GI) academicians.[8]
[9] This systems learned by iteration to place a green box on any polyp-like lesion
drawing the attention of the endoscopist to see that area more carefully. This is
done by pattern recognition by various optical and digital characteristics of the
image and prior experience of the network. The network learned to recognize a polyp-like
lesion from prior training from multiple annotated colonoscopy videos supplied by
the colonoscopist to the endoscopy company developing the GI Genius technology. This
was the first AI-assisted colonoscopy approved by U.S. Food and Drug Administration.
Subsequently, several other manufacturers developed such AI-based colonoscopy systems.
The review by Afzalpurkar et al in this issue of the journal elucidates the development,
key studies, and limitations of computer-aided polyp detection system and computer-assisted
diagnosis system for the detection and characterization of colonic polyps.[3]
Evidences supporting the role AI on diagnosis of the colorectal polyp have been reviewed
quite comprehensively in the paper by Afzalpurkar et al in this issue of the journal.[3] The authors reviewed the development, key literature, and limitations of computer-aided
polyp detection system. Since polyps differ on their malignant potential and hence,
management depending on whether these are adenomatous with or without dysplasia or
nonadenomatous, predicting histology of the polyp from the endoscopic features is
of utmost importance. The article in the current issue of the journal also elaborates
on the current status of computer-assisted diagnosis system to characterize the polyp
further.[3] The current review is therefore of considerable importance to the endoscopists involved
in care of the patients with colorectal polyps and those who are into research in
this field. There are a few other reviews published recently that summarized and meta-analyzed
the current literature on use of AI in lower GI endoscopy.[9]
[10]
[11] Though India has made phenomenal progress in computer technology both in the field
of hardware and software, publications in relation to use of AI in the field of GI
endoscopy are quite limited.[12] It is expected that the thought-provoking review by Afzalpurkar et al will stimulate
endoscopists to take up research in this important field of application of an AI-based
computer technology to endoscopic detection of GI lesions.
