Page 63 - Read Online

Ganesan et al. Art Int Surg 2024;4:364-75 https://dx.doi.org/10.20517/ais.2024.68 Page 366

RESULTS AND DISCUSSION
AI in wound diagnosis
Wound diagnosis is challenging yet crucial to treatment. Chronic wounds most commonly include diabetic
foot ulcers, pressure injuries, arterial ulcers, and venous insufficiency ulcers. Other less common disease
states are often present where diagnoses are not obvious, such as pyoderma gangrenosum and inflammatory
ulcerations. Treating the wound commonly involves treating the underlying condition, thus making proper
diagnosis critical and the first step in treatment [Figure 1].

Experienced clinicians who have the training to properly diagnose chronic wounds are limited. For medical
students and physicians, formal wound care training is sparse. A survey of fifty American medical schools
reported that the average number of educational hours spent on the physiology of tissue injury and wound
healing over all four years of medical school was just 4.7 h . More than 47% of surveyed nurses in an
[8]
inpatient setting stated that they “did not accept wound care as a nursing task”, and more than half of the
nurses failed to provide wound care discharge education . Because of the lack of standardized education on
[9]
wound care, knowledge is often picked up through practitioner experience, creating a varied knowledge
base. The lack of standardized education has also led to the creation of organizations working to address
this issue. The Wound, Ostomy, and Continence Nurses Society is one of those organizations that provides
standardized education to nurses to help fill the need for trained wound providers. AI can similarly help
standardize and expand that base with an unlimited number of “experiences”, or data. For example, by
inputting hundreds of images of different wounds into a database, AI can examine a new wound’s image,
“compare” it to the ones in the database, and report information about the new wound.

Several researchers have utilized AI to differentiate between challenging-to-diagnose wounds [Table 1]. For
example, pyoderma gangrenosum is easily misdiagnosed as cellulitis, diabetic foot ulcers, abscesses, and
other processes. The misdiagnosis of pyoderma gangrenosum can expose patients to risks that are
associated with its treatment and delay care for other causes of ulceration . It can lead to prolonged
[10]
hospitalization, unnecessary procedures, and increased medical costs for the hospital and the patient.
Birkner et al., however, developed a deep convolutional neural network to differentiate pyoderma
gangrenosum from conventional leg ulcers with a higher sensitivity than trained dermatologists . This
[11]
technology can help prevent misdiagnosis.

Similarly, Hüsers et al. studied image detection and classification algorithms for venous leg ulcers and
diabetic foot ulcers, and their algorithms of the YoloV5 (“You-Only-Look-Once”) family resulted in a high
precision (0.94) . With such high precision, this technology could serve as a tool for double-checking
[12]
physicians, enhancing their confidence that they are accurately diagnosing and treating patients.

Several deep learning tools, involving superpixel segmentation and a convolutional neural network, have
been created to classify pressure and diabetic wound images with higher accuracy than what had been done
in the past [13-16] . One model, “Alexnet architecture”, attained about 99% accuracy, 99% sensitivity, and 99%
specificity . Such high values are necessary for monitoring the progress of healing. Mohammed et al. used
[16]
an AI digital application to capture quality wound images and calculate wound surface area faster than
clinic staff using a standard digital camera, saving about two minutes on each wound assessment .
[17]
AI is quick and efficient, facilitating noncontact optical assessment of a patient’s wound, which can
potentially reduce pain and risk of infection. It also allows non-providers to assess wounds, which is crucial
as they often require daily assistance from family members, friends, or nonmedical caretakers. For example,
Lau et al. developed a smartphone application to perform real-time detection and staging classification of