Hybrid SqueezeNet-LSTM framework with advanced SegNet segmentation for automated skin disease detection

Anantha Reddy Dasari; Saroj Shambharkar; Jaykumar Lachure; Vijay Kumar Damera; Sagar Lachure

Article Details

Hybrid SqueezeNet-LSTM framework with advanced SegNet segmentation for automated skin disease detection

Authors:

Anantha Reddy Dasari | Saroj Shambharkar | Jaykumar Lachure | Vijay Kumar Damera | Sagar Lachure

Publisher:

Hacettepe University

Main Topics:

Deep Learning

Journal:

Hacettepe Journal of Mathematics and Statistics Vol.54 Issue 4, pp.1657-1687

Document Type:

Article

Publication Date:

8/2025

Document Language:

English

Document ID (Org No.):

2001356978

pISSN:

2651-477X

Keywords:

Abstract:

 Skin diseases such as pyoderma, scabies, and fungal infections remain a pressing public health concern in India due to poor hygiene, overcrowding, and limited access to care. To address these challenges, this study introduces a novel deep learning framework, the SqueezeNet-Modified Long-Short-Term Memory model, for the automated detection of skin diseases. The system incorporates four core phases: preprocessing via Gaussian filtering to reduce image noise, segmentation using a Modified SegNet enhanced with a Beta-softmax activation for precise lesion isolation, hybrid feature extraction combining shape, texture, colour, d Local Gradient Triangular Patter, and deep features, and robust classification through the SqueezeNet-Modified Long Short-Term Memory model integrated with Multi-Region Window pooling pooling and Focal-log-cosh loss. The innovative Beta-softmax function and Multi-Region Window pooling strategies enhance feature prioritization and classification accuracy. Evaluation in two data sets, one with 1,414 images of vitiligo and psoriasis, and another with 61 samples in four skin conditions, demonstrates superior performance accuracy: 0.958, sensitivity: 0.953, specificity: 0.948, F-measure: 0.950 over baseline models such as long short-term memory and novel segmented neural networks. This framework provides a scalable solution for dermatological diagnostics in low-resource settings, with future enhancements targeting the expansion to transformer-based approaches and larger clinical data sets.

 Skin diseases such as pyoderma, scabies, and fungal infections remain a pressing public health concern in India due to poor hygiene, overcrowding, and limited access to care. To address these challenges, this study introduces a novel deep learning framework, the SqueezeNet-Modified Long-Short-Term Memory model, for the automated detection of skin diseases. The system incorporates four core phases: preprocessing via Gaussian filtering to reduce image noise, segmentation using a Modified SegNet enhanced with a Beta-softmax activation for precise lesion isolation, hybrid feature extraction combining shape, texture, colour, d Local Gradient Triangular Patter, and deep features, and robust classification through the SqueezeNet-Modified Long Short-Term Memory model integrated with Multi-Region Window pooling pooling and Focal-log-cosh loss. The innovative Beta-softmax function and Multi-Region Window pooling strategies enhance feature prioritization and classification accuracy. Evaluation in two data sets, one with 1,414 images of vitiligo and psoriasis, and another with 61 samples in four skin conditions, demonstrates superior performance accuracy: 0.958, sensitivity: 0.953, specificity: 0.948, F-measure: 0.950 over baseline models such as long short-term memory and novel segmented neural networks. This framework provides a scalable solution for dermatological diagnostics in low-resource settings, with future enhancements targeting the expansion to transformer-based approaches and larger clinical data sets.

 Skin diseases such as pyoderma, scabies, and fungal infections remain a pressing public health concern in India due to poor hygiene, overcrowding, and limited access to care. To address these challenges, this study introduces a novel deep learning framework, the SqueezeNet-Modified Long-Short-Term Memory model, for the automated detection of skin diseases. The system incorporates four core phases: preprocessing via Gaussian filtering to reduce image noise, segmentation using a Modified SegNet enhanced with a Beta-softmax activation for precise lesion isolation, hybrid feature extraction combining shape, texture, colour, d Local Gradient Triangular Patter, and deep features, and robust classification through the SqueezeNet-Modified Long Short-Term Memory model integrated with Multi-Region Window pooling pooling and Focal-log-cosh loss. The innovative Beta-softmax function and Multi-Region Window pooling strategies enhance feature prioritization and classification accuracy. Evaluation in two data sets, one with 1,414 images of vitiligo and psoriasis, and another with 61 samples in four skin conditions, demonstrates superior performance accuracy: 0.958, sensitivity: 0.953, specificity: 0.948, F-measure: 0.950 over baseline models such as long short-term memory and novel segmented neural networks. This framework provides a scalable solution for dermatological diagnostics in low-resource settings, with future enhancements targeting the expansion to transformer-based approaches and larger clinical data sets.

 Skin diseases such as pyoderma, scabies, and fungal infections remain a pressing public health concern in India due to poor hygiene, overcrowding, and limited access to care. To address these challenges, this study introduces a novel deep learning framework, the SqueezeNet-Modified Long-Short-Term Memory model, for the automated detection of skin diseases. The system incorporates four core phases: preprocessing via Gaussian filtering to reduce image noise, segmentation using a Modified SegNet enhanced with a Beta-softmax activation for precise lesion isolation, hybrid feature extraction combining shape, texture, colour, d Local Gradient Triangular Patter, and deep features, and robust classification through the SqueezeNet-Modified Long Short-Term Memory model integrated with Multi-Region Window pooling pooling and Focal-log-cosh loss. The innovative Beta-softmax function and Multi-Region Window pooling strategies enhance feature prioritization and classification accuracy. Evaluation in two data sets, one with 1,414 images of vitiligo and psoriasis, and another with 61 samples in four skin conditions, demonstrates superior performance accuracy: 0.958, sensitivity: 0.953, specificity: 0.948, F-measure: 0.950 over baseline models such as long short-term memory and novel segmented neural networks. This framework provides a scalable solution for dermatological diagnostics in low-resource settings, with future enhancements targeting the expansion to transformer-based approaches and larger clinical data sets.

Article Details

Hybrid SqueezeNet-LSTM framework with advanced SegNet segmentation for automated skin disease detection

Authors:

Anantha Reddy Dasari | Saroj Shambharkar | Jaykumar Lachure | Vijay Kumar Damera | Sagar Lachure

Publisher:

Hacettepe University

Main Topics:

Deep Learning

Journal:

Hacettepe Journal of Mathematics and Statistics Vol.54 Issue 4, pp.1657-1687

Document Type:

Article

Publication Date:

8/2025

Document Language:

English

Document ID (Org No.):

2001356978

pISSN:

2651-477X

Skin diseases

deep learning

SqueezeNet-Modified LSTM

automated diagnosis

image segmentation

Issues

Article Details

Hybrid SqueezeNet-LSTM framework with advanced SegNet segmentation for automated skin disease detection

Authors: Anantha Reddy Dasari | Saroj Shambharkar | Jaykumar Lachure | Vijay Kumar Damera | Sagar Lachure

Publisher: Hacettepe University

Main Topics: Deep Learning

Journal: Hacettepe Journal of Mathematics and Statistics Vol.54 Issue 4, pp.1657-1687

Document Type: Article

Publication Date: 8/2025

Document Language: English

Document ID (Org No.): 2001356978

pISSN: 2651-477X

Skin diseases

deep learning

SqueezeNet-Modified LSTM

automated diagnosis

image segmentation

Issues

Authors:

Anantha Reddy Dasari | Saroj Shambharkar | Jaykumar Lachure | Vijay Kumar Damera | Sagar Lachure

Publisher:

Hacettepe University

Main Topics:

Deep Learning

Journal:

Hacettepe Journal of Mathematics and Statistics Vol.54 Issue 4, pp.1657-1687

Document Type:

Article

Publication Date:

8/2025

Document Language:

English

Document ID (Org No.):

2001356978

pISSN:

2651-477X