<p>Hyperparameter optimization (HPO) is essential for deep learning in medical image classification, yet standard metaheuristics such as Manta Ray Foraging Optimization (MRFO) often suffer from premature convergence in high-dimensional search spaces. To address these limitations, an enhanced variant, MRFO-LF, was proposed by incorporating Lévy flight-based exploration, adaptive step-size decay, and a hybrid stochastic–deterministic search mechanism. This work details the first application of the proposed MRFO-LF to HPO in melanoma classification, a critical task within medical image analysis. The Lévy component enables long-range perturbations, while the adaptive decay mechanism gradually narrows the search scope, and the hybrid strategy balances global versus local exploration without relying on problem-specific heuristics. Experiments were conducted on the ISIC and PH<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44443_2025_78_Article_IEq1.gif" Format="GIF" Height="10" Rendition="HTML" Resolution="72" Type="Linedraw" Width="8" /> </InlineMediaObject> <EquationSource Format="TEX">$ ^2 $</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mn>2</mn> </mmultiscripts> </math></EquationSource> </InlineEquation> dermoscopic datasets using DenseNet121, InceptionV3, and VGG19. MRFO-LF attained peak validation accuracies of 99.49% (ISIC) and 100.00% (PH<InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44443_2025_78_Article_IEq1.gif" Format="GIF" Height="10" Rendition="HTML" Resolution="72" Type="Linedraw" Width="8" /> </InlineMediaObject> <EquationSource Format="TEX">$ ^2 $</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mn>2</mn> </mmultiscripts> </math></EquationSource> </InlineEquation>) for DenseNet121, with corresponding validation losses of 0.3580 and 0.0015. When compared to MRFO, PSO, and GA, the proposed method improved ISIC accuracy by 0.40%, reduced PH<InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44443_2025_78_Article_IEq1.gif" Format="GIF" Height="10" Rendition="HTML" Resolution="72" Type="Linedraw" Width="8" /> </InlineMediaObject> <EquationSource Format="TEX">$ ^2 $</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mn>2</mn> </mmultiscripts> </math></EquationSource> </InlineEquation> loss by over 95%, and converged up to 30% faster. Statistical significance was confirmed through ANOVA and paired t-tests (<InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44443_2025_78_Article_IEq4.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="62" /> </InlineMediaObject> <EquationSource Format="TEX">$ p < 0.05 $</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>p</mi> <mo><</mo> <mn>0.05</mn> </mrow> </math></EquationSource> </InlineEquation>). These results position MRFO-LF as a reliable and efficient optimizer for complex hyperparameter tuning in medical image classification.</p>

Adaptive hybrid hyperparameter optimization with MRFO and Lévy flight for accurate melanoma classification

期刊论文
出版模式：

开放获取
发表日期：

2025年6月12日

查看全文

Shamsuddeen Adamu,
Hitham Alhussian,
Said Jadid Abdulkadir,
Ayed Alwadin,
Sallam O. F. Khairy,
Hussaini Mamman,
Shamsu Abdullahi,
Saidu Yahaya,
Aliyu Garba,
Dahiru Adamu Aliyu,
Muhammad Muntasir Yakubu,
Daniel Tonye Oyefidein

摘要

Hyperparameter optimization (HPO) is essential for deep learning in medical image classification, yet standard metaheuristics such as Manta Ray Foraging Optimization (MRFO) often suffer from premature convergence in high-dimensional search spaces. To address these limitations, an enhanced variant, MRFO-LF, was proposed by incorporating Lévy flight-based exploration, adaptive step-size decay, and a hybrid stochastic–deterministic search mechanism. This work details the first application of the proposed MRFO-LF to HPO in melanoma classification, a critical task within medical image analysis. The Lévy component enables long-range perturbations, while the adaptive decay mechanism gradually narrows the search scope, and the hybrid strategy balances global versus local exploration without relying on problem-specific heuristics. Experiments were conducted on the ISIC and PH $ ^2 $ $^{2}$ dermoscopic datasets using DenseNet121, InceptionV3, and VGG19. MRFO-LF attained peak validation accuracies of 99.49% (ISIC) and 100.00% (PH $ ^2 $ $^{2}$ ) for DenseNet121, with corresponding validation losses of 0.3580 and 0.0015. When compared to MRFO, PSO, and GA, the proposed method improved ISIC accuracy by 0.40%, reduced PH $ ^2 $ $^{2}$ loss by over 95%, and converged up to 30% faster. Statistical significance was confirmed through ANOVA and paired t-tests ( $ p < 0.05 $ $p < 0.05$ ). These results position MRFO-LF as a reliable and efficient optimizer for complex hyperparameter tuning in medical image classification.