Source Datasets Research Articles

Multilayer meta-matching: Translating phenotypic prediction models from multiple datasets to small data

Abstract Resting-state functional connectivity (RSFC) is widely used to predict phenotypic traits in individuals. Large sample sizes can significantly improve prediction accuracies. However, for studies of certain clinical populations or focused neuroscience inquiries, small-scale datasets often remain a necessity. We have previously proposed a “meta-matching” approach to translate prediction models from large datasets to predict new phenotypes in small datasets. We demonstrated a large improvement over classical kernel ridge regression (KRR) when translating models from a single source dataset (UK Biobank) to the Human Connectome Project Young Adults (HCP-YA) dataset. In the current study, we propose two meta-matching variants (“meta-matching with dataset stacking” and “multilayer meta-matching”) to translate models from multiple source datasets across disparate sample sizes to predict new phenotypes in small target datasets. We evaluate both approaches by translating models trained from five source datasets (with sample sizes ranging from 862 participants to 36,834 participants) to predict phenotypes in the HCP-YA and HCP-Aging datasets. We find that multilayer meta-matching modestly outperforms meta-matching with dataset stacking. Both meta-matching variants perform better than the original “meta-matching with stacking” approach trained only on the UK Biobank. All meta-matching variants outperform classical KRR and transfer learning by a large margin. In fact, KRR is better than classical transfer learning when less than 50 participants are available for finetuning, suggesting the difficulty of classical transfer learning in the very small sample regime. The multilayer meta-matching model is publicly available at https://github.com/ThomasYeoLab/Meta_matching_models/tree/main/rs-fMRI/v2.0.

Improved random forest model for smoke detection and early warning based on sparrow search algorithm

Smoke detection technology is of great significance in the field of fire safety, aiming at the timely detection of smoke released from fires or other combustion events to safeguard people's lives and properties. In this paper, a smart smoke alarm device is developed based on AI algorithm using UCL open source dataset, which is processed by AI recognition after collecting data through IoT to achieve accurate judgement of smoke situation. The dataset is divided by 6:4 ratio, and the training set confusion matrix shows that all 1159 smoke alarm tests are correctly predicted, of which 445 times predict no smoke and 714 times predict smoke, with an accuracy rate of 100%. The test set confusion matrix shows that only 1 out of 1158 tests was incorrectly predicted, of which 184 predicted no smoke, 312 predicted smoke, and 1 was misjudged as no smoke even though it should have been smoke, with an accuracy rate of 99.7%. The successful application of this technology provides a reliable guarantee for fire warning and demonstrates the great potential of AI in the field of security.

Unsupervised shape-and-texture-based generative adversarial tuning of pre-trained networks for carotid segmentation from 3D ultrasound images.

Vessel-wall volume and localized three-dimensional ultrasound (3DUS) metrics are sensitive to the change of carotid atherosclerosis in response to medical/dietary interventions. Manual segmentation of the media-adventitia boundary (MAB) and lumen-intima boundary (LIB) required to obtain these metrics is time-consuming and prone to observer variability. Although supervised deep-learning segmentation models have been proposed, training of these models requires a sizeable manually segmented training set, making larger clinical studiesprohibitive. We aim to develop a method to optimize pre-trained segmentation models without requiring manual segmentation to supervise the fine-tuningprocess. We developed an adversarial framework called the unsupervised shape-and-texture generative adversarial network (USTGAN) to fine-tune a convolutional neural network (CNN) pre-trained on a source dataset for accurate segmentation of a target dataset. The network integrates a novel texture-based discriminator with a shape-based discriminator, which together provide feedback for the CNN to segment the target images in a similar way as the source images. The texture-based discriminator increases the accuracy of the CNN in locating the artery, thereby lowering the number of failed segmentations. Failed segmentation was further reduced by a self-checking mechanism to flag longitudinal discontinuity of the artery and by self-correction strategies involving surface interpolation followed by a case-specific tuning of the CNN. The U-Net was pre-trained by the source dataset involving 224 3DUS volumes with 136, 44, and 44 volumes in the training, validation and testing sets. The training of USTGAN involved the same training group of 136 volumes in the source dataset and 533 volumes in the target dataset. No segmented boundaries for the target cohort were available for training USTGAN. The validation and testing of USTGAN involved 118 and 104 volumes from the target cohort, respectively. The segmentation accuracy was quantified by Dice Similarity Coefficient (DSC), and incorrect localization rate (ILR). Tukey's Honestly Significant Difference multiple comparison test was employed to quantify the difference of DSCs between models and settings, where was considered statisticallysignificant. USTGAN attained a DSC of % in LIB and % in MAB, improving from the baseline performance of % in LIB (p ) and % in MAB (p ). Our approach outperformed six state-of-the-art domain-adaptation models (MAB: , LIB: ). The proposed USTGAN also had the lowest ILR among the methods compared (LIB: 2.5%, MAB: 1.7%). Our framework improves segmentation generalizability, thereby facilitating efficient carotid disease monitoring in multicenter trials and in clinics with less expertise in 3DUSimaging.

Spectral Decomposition and Transformation for Cross-domain Few-shot Learning

Cross-domain few-shot Learning (CDFSL) is proposed to first pre-train deep models on a source domain dataset where sufficient data is available, and then generalize models to target domains to learn from only limited data. However, the gap between the source and target domains greatly hampers the generalization and target-domain few-shot finetuning. To address this problem, we analyze the domain gap from the aspect of frequency-domain analysis. We find the domain gap could be reflected by the compositions of source-domain spectra, and the lack of compositions in the source datasets limits the generalization. Therefore, we aim to expand the coverage of spectra composition in the source datasets to help the source domain cover a larger range of possible target-domain information, to mitigate the domain gap. To achieve this goal, we propose the Spectral Decomposition and Transformation (SDT) method, which first randomly decomposes the spectrogram of the source datasets into orthogonal bases, and then randomly samples different coordinates in the space formed by these bases. We integrate the above process into a data augmentation module, and further design a two-stream network to handle augmented images and original images respectively. Experimental results show that our method achieves state-of-the-art performance in the CDFSL benchmark dataset.

Pano2Geo: An efficient and robust building height estimation model using street-view panoramas

Building height serves as a crucial parameter in characterizing urban vertical structure, which has a profound impact on urban sustainable development. The emergence of street-view data offers the opportunity to observe urban 3D scenarios from the human perspective, benefiting the estimation of building height. In this paper, we propose an efficient and robust building height estimation model, which we call the Pano2Geo model, by precisely projecting street-view panorama (SVP) coordinates to geospatial coordinates. Firstly, an SVP refinement stratagem is designed, incorporating NENO rules for observation quality assessment from four aspects: number of buildings, extent of the buildings, number of nodes, and orthogonal observations, followed by the application of the art gallery theorem to further refine the SVPs. Secondly, the Pano2Geo model is constructed, which provides a pixel-level projection transformation from SVP coordinates to 3D geospatial coordinates for locating the height features of buildings in the SVP. Finally, the valid building height feature points in the SVP are extracted based on a slope mutation test, and the 3D geospatial coordinates of the building height feature points are projected using the Pano2Geo model, so as to obtain the building height. The proposed model was evaluated in the city of Wuhan in China, and the results indicate that the Pano2Geo model can accurately estimate building height, with an average error of 1.85 m. Furthermore, compared with three state-of-the-art methods, the Pano2Geo model shows superior performance, with only 10.2 % of buildings have absolute errors exceeding 2 m, compared to the Map-image-based (27.2 %), Corner-based (16.8 %), and Single-view-based (13.9 %) height estimation methods. The SVP refinement method achieves optimal observation quality with less than 50 % of existing SVPs, leading to highly efficient building height estimation, particularly in areas of a high building density. Moreover, the Pano2Geo model exhibits robustness in building height estimation, maintaining errors within 2 m even as building shape complexity and occlusion degree increase within the SVP. Our source dataset and code are available at https://github.com/Giser317/Pano2Geo.git.

Tensorized Incomplete Multi-view Kernel Subspace Clustering

Recently considerable advances have been achieved in the incomplete multi-view clustering (IMC) research. However, the current IMC works are often faced with three challenging issues. First, they mostly lack the ability to recover the nonlinear subspace structures in the multiple kernel spaces. Second, they usually neglect the high-order relationship in multiple representations. Third, they often have two or even more hyper-parameters and may not be practical for some real-world applications. To tackle these issues, we present a Tensorized Incomplete Multi-view Kernel Subspace Clustering (TIMKSC) approach. Specifically, by incorporating the kernel learning technique into an incomplete subspace clustering framework, our approach can robustly explore the latent subspace structure hidden in multiple views. Furthermore, we impute the incomplete kernel matrices and learn the low-rank tensor representations in a mutual enhancement manner. Notably, our approach can discover the underlying relationship among the observed and missing samples while capturing the high-order correlation to assist subspace clustering. To solve the proposed optimization model, we design a three-step algorithm to efficiently minimize the unified objective function, which only involves one hyper-parameter that requires tuning. Experiments on various benchmark datasets demonstrate the superiority of our approach. The source code and datasets are available at: https://www.researchgate.net/publication/381828300_TIMKSC_20240629.

IoV-BCFL: An intrusion detection method for IoV based on blockchain and federated learning

In recent years, Internet of Vehicles (IoV) is in a booming stage. But at the same time, the methods of attack against IoV such as Denial of Service (DoS) and deception are great threats to personal and social security. Traditional IoV intrusion detection usually adopts a centralized detection model, which has the disadvantages of untimely detection results and is difficult to protect vehicle privacy in practical applications. Meanwhile, centralized computation requires a large amount of vehicle data transmission, which overloads the wireless bandwidth. Combined the distributed computing resources of Federated Learning (FL) and the decentralized features of blockchain, an IoV intrusion detection framework named IoV-BCFL is proposed, which is capable of distributed intrusion detection and reliable logging with privacy protection. FL is used for distributing training on vehicle nodes and aggregating the training models at Road Side Unit (RSU) to reduce data transmission, protect the privacy of training data, and ensure the security of the model. A blockchain-based intrusion logging mechanism is presented, which enhances vehicle privacy protection through Rivest-Shamir-Adleman (RSA) algorithm encryption and uses Inter Planetary File System (IPFS) to store the intrusion logs. The intrusion behavior can be faithfully recorded by logging smart contract after detecting the intrusion, which can be used to track intruders, analyze security vulnerabilities, and collect evidence. Experiments based on different open source datasets show that FL achieves a high detection rates on intrusion data and effectively reduce the communication overhead, the smart contract performs well on evaluation indicators such as sending rate, latency, and throughput.

Dense Continuous-Time Optical Flow From Event Cameras.

We present a method for estimating dense continuous-time optical flow from event data. Traditional dense optical flow methods compute the pixel displacement between two images. Due to missing information, these approaches cannot recover the pixel trajectories in the blind time between two images. In this work, we show that it is possible to compute per-pixel, continuous-time optical flow using events from an event camera. Events provide temporally fine-grained information about movement in pixel space due to their asynchronous nature and microsecond response time. We leverage these benefits to predict pixel trajectories densely in continuous time via parameterized Bézier curves. To achieve this, we build a neural network with strong inductive biases for this task: First, we build multiple sequential correlation volumes in time using event data. Second, we use Bézier curves to index these correlation volumes at multiple timestamps along the trajectory. Third, we use the retrieved correlation to update the Bézier curve representations iteratively. Our method can optionally include image pairs to boost performance further. To the best of our knowledge, our model is the first method that can regress dense pixel trajectories from event data. To train and evaluate our model, we introduce a synthetic dataset (MultiFlow) that features moving objects and ground truth trajectories for every pixel. Our quantitative experiments not only suggest that our method successfully predicts pixel trajectories in continuous time but also that it is competitive in the traditional two-view pixel displacement metric on MultiFlow and DSEC-Flow. Open source code and datasets are released to the public.

CMM: A CNN-MLP Model for COVID-19 Lesion Segmentation and Severity Grading.

In this paper, a CNN-MLP model (CMM) is proposed for COVID-19 lesion segmentation and severity grading in CT images. The CMM starts by lung segmentation using UNet, and then segmenting the lesion from the lung region using a multi-scale deep supervised UNet (MDS-UNet), finally implementing the severity grading by a multi-layer preceptor (MLP). In MDS-UNet, shape prior information is fused with the input CT image to reduce the searching space of the potential segmentation outputs. The multi-scale input compensates for the loss of edge contour information in convolution operations. In order to enhance the learning of multiscale features, the multi-scale deep supervision extracts supervision signals from different upsampling points on the network. In addition, it is empirical that the lesion which has a whiter and denser appearance tends to be more severe in the COVID-19 CT image. So, the weighted mean gray-scale value (WMG) is proposed to depict this appearance, and together with the lung and lesion area to serve as input features for the severity grading in MLP. To improve the precision of lesion segmentation, a label refinement method based on the Frangi vessel filter is also proposed. Comparative experiments on COVID-19 public datasets show that our proposed CMM achieves high accuracy on COVID-19 lesion segmentation and severity grading.

Knee Angle Estimation from Surface EMG during Walking Using Attention-Based Deep Recurrent Neural Networks: Feasibility and Initial Demonstration in Cerebral Palsy.

Accurately estimating knee joint angle during walking from surface electromyography (sEMG) signals can enable more natural control of wearable robotics like exoskeletons. However, challenges exist due to variability across individuals and sessions. This study evaluates an attention-based deep recurrent neural network combining gated recurrent units (GRUs) and an attention mechanism (AM) for knee angle estimation. Three experiments were conducted. First, the GRU-AM model was tested on four healthy adolescents, demonstrating improved estimation compared to GRU alone. A sensitivity analysis revealed that the key contributing muscles were the knee flexor and extensors, highlighting the ability of the AM to focus on the most salient inputs. Second, transfer learning was shown by pretraining the model on an open source dataset before additional training and testing on the four adolescents. Third, the model was progressively adapted over three sessions for one child with cerebral palsy (CP). The GRU-AM model demonstrated robust knee angle estimation across participants with healthy participants (mean RMSE 7 degrees) and participants with CP (RMSE 37 degrees). Further, estimation accuracy improved by 14 degrees on average across successive sessions of walking in the child with CP. These results demonstrate the feasibility of using attention-based deep networks for joint angle estimation in adolescents and clinical populations and support their further development for deployment in wearable robotics.

Semantic details in autobiographical memory narratives increase with age among younger adults

ABSTRACT It is well documented that older adults, compared to younger adults, produce fewer episodic details and more semantic details when recalling autobiographical memories. However, group comparisons have provided limited insight into the trajectories of detail generation across the lifespan. Utilising an open source dataset [Clark, I. A., & Maguire, E. A. (2023). Release of cognitive and multimodal MRI data including real-world tasks and hippocampal subfield segmentations. Scientific Data, 10(1), 1–29. https://doi.org/10.1038/s41597-022-01899-x], we examined how episodic and semantic detail generation varied with age among 194 younger adults, ages 20-41. We tested whether age differences were mediated by hippocampal subfield volumes and MTL resting-state functional connectivity. Results indicated that semantic details increased with age, while episodic details remained stable. We observed age differences in hippocampal subfield volumes and MTL connectivity, but these measures did not mediate age effects on semantic detail. Based on these and prior findings [Matijevic, S., Andrews-Hanna, J. R., Wank, A. A., Ryan, L., & Grilli, M. D. (2022). Individual differences in the relationship between episodic detail generation and resting state functional connectivity vary with age. Neuropsychologia, 166, 108138. https://doi.org/10.1016/j.neuropsychologia.2021.108138], we suggest a model of diverging episodic and semantic detail generation trajectories across the adult lifespan.

MDTGAN: Multi domain generative adversarial transfer learning network for traffic data imputation

Accurate, real-time, and efficient traffic data, crucial for intelligent transportation systems, which is often disrupted in the real world due to the influence of weather, interference, and equipment failures. Generative Adversarial Networks (GANs), have achieved significant results in image restoration, providing insights for traffic data imputation. Recent research has focused on developing more effective and reasonable model by learning transferable common knowledge from different cities or different scenarios, which is also an excellent idea for improving the data imputation performance. In light of, we contrive to design a GANs based transferable traffic data imputation model, namely Multi Domain Generative Adversarial Transfer Learning Network (MDTGAN). Firstly, the model consists of two stages. In pre-training stage, the model utilizes the source domain datasets (Similar cities road network datasets) to update parameters and transferred them. Then, the model parameters are optimized using the target domain dataset (Target city road network dataset) in fine-tuning stage to avoid the issue of insufficient samples caused by data missing. Secondly, to adapt the model to different dataset topologies, MDTGAN models the node-level data by taking node time series as input, enabling the model to autonomously learn the prevalent spatiotemporal correlations inherent in nodes. Additionally, we introduce a domain discriminator module that guides the spatial encoder in learning domain-invariant features of network node spatial information, enhancing the model generalization ability. The experiments conducted on three publicly available datasets, in which one dataset is regarded as the target dataset, while the others serve as source datasets. The experimental results demonstrate that the MDTGAN model consistently outperforms the baseline models. Specifically, the MDTGAN model can transfer valuable knowledge from the source domain datasets to improve data imputation performance on the target domain dataset, providing practical significance for cities lacking historical traffic data.

MDGRL: Multi-dimensional graph rule learning

Knowledge graph completion is an advanced artificial intelligence (AI) methodology that enables the systematic organization and structuring of data. It can significantly enhance the digital economy by facilitating more accurate and appropriate decision-making processes. Logical rule algorithms, known for their interpretability, have attracted significant attention in the field of explainable artificial intelligence. To harness the interpretability benefits of logical rules, we propose a cross-level position constraint template based on Graph Path Feature Learning (GPFL) and introduce an optimized termination policy for rule generation. To improve prediction performance, we focus significant emphasis on the rule evaluation stage, specifically on how to effectively learn the rules. To this end, we propose a Multi-Dimensional Graph Rule Learning (MDGRL) that calculates features from different dimensions to represent the reasoning ability of the rule. Feature I presents metrics for assessing the similarity into rule’s structure. Feature II proposes calculating the path transfer probability to represent the rule reasoning ability from the path reasoning perspective. Moreover, Feature III incorporates embedding distance with the constraint template to represent the rule reasoning ability in a low-dimensional vector space. Lastly, we assign weights to different dimensional features based on their performance and integrate them into MDGRL. Our experiments demonstrate effectiveness across various datasets, showcasing time-saving benefits in rule generation and improved prediction performance in rule evaluation. The source code and dataset for the MDGRL algorithm are available at https://github.com/csjywu1/MDGRL.

Towards episode rules with non-overlapping frequency and targeted mining

Data mining has become a popular task due to the explosion of rich data. Frequent episode mining (FEM) is an effective technique for extracting valuable and crucial information from event sequences, playing a significant role in various fields such as market basket analysis, association analysis, and management sciences. A variety of FEM algorithms discover frequent episode rules using the frequency function and anti-monotony. Non-overlapping frequency works well for mining episode rules because it cuts down on unnecessary computations and still follows anti-monotonicity. Leveraging this technique, we propose the NONEPI+ algorithm for discovering rules that were previously overlooked. However, there are no algorithms that can be used to discover episode rules that contain target query rules. To fill the research gap in episode rule mining algorithms, we further propose a novel algorithm called TaER for targeted mining of episode rules, i.e., a set of rules containing specific query rules. TaER can successfully discover the episode rules that the users are more specifically interested in. It can provide direction for prediction tasks in many aspects, such as weather observation, network intrusion, e-commerce, and financial behavior prediction. All source code and datasets are available on GitHub https://github.com/DSI-Lab1/TaER.

From “At Risk” to “Dangerous” People. The Roma Between Development Policies and Practices of Expulsion

Over more than three decades since the fall of state socialism in Central and Eastern Europe, the position of the Roma in the region is rather ambivalent. Numerous European institutions, inter-governmental development agencies, national governments and European and national/local non-governmental organizations are involved in a pan-European effort to turn the tide for the Roma. The paper argues that although there are numerous efforts to implement development policies and programs to bridge the gap between Roma and non-Roma, there is still a gap between the discourses and inclusion policies developed at the European and national level and those at the sub-national level. There is a tendency to move away from the framing of Roma as an “at risk” group to seeing them as “dangerous people”, who pose threats to the security of others. The paper uses two sets of data sources. First, it employs a secondary analysis of three recent databases: – SocioRoMap – Sociographic mapping of Roma communities in Romania for community-level monitoring of changes regarding Roma integration (2017) and two surveys conducted by the European Union Agency for Fundamental Rights (2011 and 2016). The second data source is based on semi-structured interviews carried out in several research projects that the author has participated in over the last few years. In conclusion, I argue that the two political projects of institutional developmentalism and securitization of the lifestyle of the poorest Roma can coexist, but the abandonment of the former can lead to complex social problems in the future.

The Global Historical Climatology Network Monthly Precipitation Dataset, Version 4

The Global Historical Climatology Network (GHCN) monthly precipitation dataset contains historical time series for thousands of land surface stations worldwide. Initially released in 1992 and revised in 1998, the dataset has been employed in a variety of applications over the past three decades, including operational monitoring, applied research, and international assessments. This paper describes the data and methods used to compile the latest edition (version 4), which has three major enhancements. The first enhancement is to the station network, which increased in size by a factor of five due to the inclusion of dozens of new source datasets, most notably GHCN Daily (GHCNd). The second improvement is the application of a rule-based algorithm to compare and merge records representing the same location. The third enhancement is to the quality assurance approach, now consisting of 18 new checks based on GHCNd and other operational systems. Updated monthly, the resulting dataset consists of time series of monthly precipitation totals at more than 120,000 worldwide stations, including more than 33,000 active observing sites.

An industrial heat source dataset based on remotely sensed active fire/hotspot detection in China from 2012 to 2021

An industrial heat source dataset based on remotely sensed active fire/hotspot detection in China from 2012 to 2021

Feature selection based on contradictory state sequence for multi-scale interval valued decision table

In the context of selecting features for multi-scale interval valued decision table (MIVDT), conventional research approaches encounter difficulties including elevated data complexity, diminished computational efficiency, and limited model generalization capacity. To overcome these difficulties, feature selection methods based on contradictory state sequence (CSS) and fuzzy contradictory state sequence (FCSS) are proposed in this paper. Initially, MIVDT is established. According to the characteristics of interval value, a more thorough and impartial metric for assessing the inclusion degree is suggested, along with similarity relation based on the metric. Subsequently, the introduction of the first contradictory object allows for the delineation of contradictory state and fuzzy contradictory state, which serve to characterize the consistency of MIVDT. These two concepts are proposed to enhance the accuracy and efficiency of capturing key information in intricate data. Additionally, rapid feature selection algorithms are suggested, which rely on CSS and FCSS. In contrast to conventional feature selection approaches, the algorithms introduced in this study demonstrate superior computational efficiency and enhanced generalization capabilities when confronted with intricate datasets. In twelve open source datasets, the upper limit for the quantity of objects is 110341, and the average accuracy values of experiments under two parameter combinations are 99.54% and 97.08% respectively. The results of experiments demonstrate that the algorithms introduced are capable of efficiently identifying a novel feature subset from intricate datasets within a shorter timeframe.

Towards the generalization of time series classification: A feature-level style transfer and multi-source transfer learning perspective

Transfer learning-based methods hold promise for enhancing classification task performance. However, a transfer learning mechanism for hard-to-classify time series classification tasks caused by limited samples in training set is still to be accomplished. Drawing inspiration from domain adaptation and style transfer methods in computer vision, we aim to address sample scarcity and enhance classifier generalization by leveraging abundant unrelated time series datasets. In this paper, a transfer learning mechanism called feature-level style transfer, where the feature alignment is conducted in a domain adaptation-inspired manner, has been proposed to target initially hard-to-classify time series tasks with limited training samples. The proposed mechanism eliminates constraints on source dataset selection, enabling the utilization of weakly-related or unrelated datasets to enhance classifier performance. Furthermore, a voting mechanism has been formulated to achieve more accurate predictions by taking transferable information learnt from different source domains into comprehensive consideration.

Stitched vision transformer for age-related macular degeneration detection using retinal optical coherence tomography images.

Age-related macular degeneration (AMD) is an eye disease that leads to the deterioration of the central vision area of the eye and can gradually result in vision loss in elderly individuals. Early identification of this disease can significantly impact patient treatment outcomes. Furthermore, given the increasing elderly population globally, the importance of automated methods for rapidly monitoring at-risk individuals and accurately diagnosing AMD is growing daily. One standard method for diagnosing AMD is using optical coherence tomography (OCT) images as a non-invasive imaging technology. In recent years, numerous deep neural networks have been proposed for the classification of OCT images. Utilizing pre-trained neural networks can speed up model deployment in related tasks without compromising accuracy. However, most previous methods overlook the feasibility of leveraging pre-existing trained networks to search for an optimal architecture for AMD staging on a new target dataset. In this study, our objective was to achieve an optimal architecture in the efficiency-accuracy trade-off for classifying retinal OCT images. To this end, we employed pre-trained medical vision transformer (MedViT) models. MedViT combines convolutional and transformer neural networks, explicitly designed for medical image classification. Our approach involved pre-training two distinct MedViT models on a source dataset with labels identical to those in the target dataset. This pre-training was conducted in a supervised manner. Subsequently, we evaluated the performance of the pre-trained MedViT models for classifying retinal OCT images from the target Noor Eye Hospital (NEH) dataset into the normal, drusen, and choroidal neovascularization (CNV) classes in zero-shot settings and through five-fold cross-validation. Then, we proposed a stitching approach to search for an optimal model from two MedViT family models. The proposed stitching method is an efficient architecture search algorithm known as stitchable neural networks. Stitchable neural networks create a candidate model in search space for each pair of stitchable layers by inserting a linear layer between them. A pair of stitchable layers consists of layers, each selected from one input model. While stitchable neural networks had previously been tested on more extensive and general datasets, this study demonstrated that stitching networks could also be helpful in smaller medical datasets. The results of this approach indicate that when pre-trained models were available for OCT images from another dataset, it was possible to achieve a model in 100 epochs with an accuracy of over 94.9% in classifying images from the NEH dataset. The results of this study demonstrate the efficacy of stitchable neural networks as a fine-tuning method for OCT image classification. This approach not only leads to higher accuracy but also considers architecture optimization at a reasonable computational cost.