Call us Today ! +918886268863 | [email protected]

Robust Skin Disease Classification by Distilling Deep Neural Network Ensemble for the Mobile Diagnosis of Herpes Zoster

Robust Skin Disease Classification by Distilling Deep Neural Network Ensemble for the Mobile Diagnosis of Herpes Zoster

Introduction:

Herpes zoster (HZ) is a virus-induced skin disease characterized by a painful rash accompanied by blisters. The occurrence of HZ in a lifetime is 10%_30% [1], [2]. However, if proper antiviral treatments are not provided within seventy two hours after the onset of a rash, it can progress to chronic disease with severe pain [3]. Thus, early diagnosis of HZ is crucial for a complete recovery; otherwise, it leads to severe complications. Despite its frequent occurrence and severity, most people are unaware of their risk for HZ [4]. The onset of HZ is accompanied by mild symptoms such as fever, itching, and chills and it can progress to persistent neuropathic pain. HZ decreases the quality of life, which severely affects an individual’s sleep and social activities.

Abstract:

Herpes zoster (HZ) is a common cutaneous disease affecting one out of _ve people; hence, early diagnosis of HZ is crucial as it can progress to chronic pain syndrome if antiviral treatment is not provided within 72 hr. Mobile diagnosis of HZ with the assistance of artificial intelligence can prevent neuropathic pain while reducing clinicians’ fatigue and diagnosis cost. However, the clinical images captured from daily mobile devices likely contain visual corruptions, such as motion blur and noise, which can easily mislead the automated system. Hence, this paper aims to train a robust and mobile deep neural network (DNN) that can distinguish HZ from other skin diseases using user-submitted images. To enhance robustness while retaining low computational cost, we propose a knowledge distillation from ensemble via curriculum training (KDE-CT) wherein a student network learns from a stronger teacher network progressively. We established skin diseases dataset for HZ diagnosis and evaluated the robustness against 75 types of corruption. A total of 13 different DNNs was evaluated on both clean and corrupted images.

Existing work:

In recent years, some studies have considered clinical images to classify skin lesion. Collected a large number of skin images from 16,114 cases, including both dermoscopic and clinical images for the differential diagnosis of 26 skin conditions using deep learning. Trained a DNN using 220,680 clinical images to suggest the treatment options for 134 skin disorders. Although the aforementioned models achieved promising performance and were extensively applied in the clinical domain. Utilizes a large neural network consisting of one to six Inception-V4 networks [13], and [12] uses an ensemble of four different convolutional networks.

Disadvantage:

It is difficult to incorporate these models directly into mobile devices owing to their high computational costs.

Proposed work:

We propose a knowledge distillation from ensemble via curriculum training (KDE-CT) wherein a student network learns from a stronger teacher network progressively. We established skin diseases dataset for HZ diagnosis and evaluated the robustness against 75 types of corruption. A total of 13 different DNNs was evaluated on both clean and corrupted images.

Advantage:

  • High performance

Algorithm:

System requirements:

  Software requirements:

  • Operating system           :   Windows.
  • Coding Language :   Python.

Hardware components:

System                 :   Pentium IV 2.4 GHz or intel

Hard Disk                      :   40 GB.

Floppy Drive       :   1.44 Mb.

Mouse                  :   Optical Mouse.

Ram                      :   512 Mb.

Conclusion:

This study aimed to distinguish HZ from the other skin diseases for mobile applications. We established an SD-HZ dataset comprising 413 single images and 8,345 clinical images. We also constructed an SD-HZ-C dataset with 15 types of corruption with five severity levels for each type to evaluate robustness against input visual corruption. A total of 13 different DNNs were trained on SD-HZ images and were evaluated on both clean and corrupted images. The results showed that corruption error should be considered along with accuracy when selecting an appropriate model for mobile skin disease diagnosis. In this regard, MobileNet-V3-small represents a reasonable choice considering efficiency as well as accuracy. We proposed KDE-CT to enhance the robustness of DNN by progressively changing the teacher network and verified that it can be an effective solution for improving the corruption robustness while retaining satisfactory performance on clean images.

March 19, 2022

0 responses on "Robust Skin Disease Classification by Distilling Deep Neural Network Ensemble for the Mobile Diagnosis of Herpes Zoster"

Leave a Message

Template Design © VibeThemes. All rights reserved.