WSWindra Swastika

Windra Swastika

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Windra Swastika has a total of 27 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Mask-Prior Knowledge Distillation for Efficient Binary Breast Ultrasound Classification

    by Windra Swastika, Paulus Lucky Tirma Irawan

    A peer-reviewed journal of applied informatics that covers a broad range of topics, especially dealing with the social and business aspects of informatics.

  • Open Access

    A deep learning-based mobile decision support system for optimizing stress induced fruiting in Citrus reticulata

    by Kestrilia Rega Prilianti, Windra Swastika, Handika Gunawan Putra, Meme Susilowati, Otto Endarto

    Purpose: This study proposes a deep learning-based decision support system for stress induction management in tangerine cultivation. While current techniques depend heavily on subjective visual information, we offer an objective, data-driven mechanism to determine whether a branch is likely to respond favorably to stress induction. Design/methodology/approach: Leaf images were used as input, as color reflects photosynthetic pigment composition linked to readiness for stress-induced flowering. A dataset of 516 images, acquired under controlled conditions, was used to train the model. The system predicts flowering success into three categories: low, moderate and high. Four convolutional neural networks architectures (VGG16, ResNet50, MobileNet and Inception) were evaluated using Adam and Adamax with 10-fold cross-validation. Beyond standard transfer learning, the modeling pipeline integrates task-specific fine-tuning and red-, green- and blue-based colorimetric interpretation. Performance analysis includes receiver operating characteristic and gradient-weighted class activation mapping, with metrics summarized using accuracy, precision, recall and F1-score. Model selection was based on class-wise robustness and model size. Findings: We found that architectural choice strongly affects predictive performance. The inception provides the highest accuracy-efficiency ratio (accuracy = 0.90, F1-Score = 0.89, precision = 0.92, recall = 0.88 and model size = 57 Megabytes), beating larger models and being lightweight enough for mobile application. The trained models were implemented into a mobile application prototype for real-time inference. The low flowering-success category remains difficult to detect due to its high intra-class variability. Originality/value: This decision support system provides a practical approach to use an artificial intelligence-driven flowering analysis straight in stress induction workflows for more accurate and efficient cultivation management.

  • Open Access

    Finding the Sweet Spot: A Study of Data Augmentation Intensity for Small-Scale Image Classification

    by Windra Swastika

    Finding the optimal level of data augmentation intensity remains one of the most challenging aspects of training deep learning models on small-scale datasets. While data augmentation is universally recognized as essential for preventing overfitting and improving generalization, excessive augmentation can paradoxically harm model performance by introducing too much variability in the training data. This research investigates the "sweet spot" of augmentation intensity through a comprehensive study of six distinct augmentation strategies on CIFAR-10, a representative small-scale image classification benchmark. We designed a controlled experiment comparing: No Augmentation (baseline), Basic torchvision transforms, Light Advanced albumentations, Moderate Advanced geometric-photometric combinations, Strong Advanced with noise injection and dropout, and AutoAugment Style with complex transformations. Our findings reveal a clear quadratic relationship between augmentation intensity and model performance (R² = 0.85), with peak performance achieved at moderate intensity levels (0.21 on our normalized scale). The Basic augmentation strategy wins with 79.04% validation accuracy, significantly outperforming both minimal augmentation (77.15%) and excessive augmentation (72.17%). Through statistical analysis including effect size calculations, confidence intervals, and correlation studies, we demonstrate that the "sweet spot" lies in balanced augmentation that provides regularization benefits without overwhelming the learning process. These findings offer evidence-based guidelines for practitioners working with small-scale image datasets.

  • Open Access

    Implementasi Strategi Trading Cryptocurrency Menggunakan Indikator QMI: Perspektif Sebelas Tahun pada Bitcoin (2013–2024)

    by Windra Swastika, Tarsisius Renald Suganda, Rino Tam Cahyadi

    Research aim : This study aims to develop and test the effectiveness of a Quantitative Market Indicator (QMI)-based trading strategy to optimize Bitcoin investment returns while managing risk through controlled position allocation.. Design/Methode/Approach : The research follows five stages: data preprocessing, market indicator calculation, trading strategy design, performance analysis, and visualization. The QMI integrates three key indicators—Puell Multiple, Golden Fibonacci Index, and Pi Cycle—to identify market cycles and trend reversals. The systematic trading strategy applies buying rules when QMI < 20 and selling when QMI > 80, allocating 3% of capital for each transaction. Research Finding : The QMI strategy generated 1,101 transactions (479 buys, 622 sells), achieving a 12,670.71% return with an initial capital of USD 1,000 growing to USD 127,700.32. The maximum drawdown was 32.4%, the win rate 68.5%, and the Sharpe ratio 2.1, indicating strong performance with controlled risk. Theoretical contribution/Originality : This study introduces the integration of the Fibonacci sequence and the golden ratio in market cycle detection, enhancing prediction accuracy in volatile crypto markets. Practitioner/Policy implication : The QMI model provides a structured decision-making framework for crypto investors, emphasizing disciplined trading and adaptive risk management. Research limitation : Further development should integrate market sentiment and dynamic thresholds to refine model robustness.

  • Study of EMG-based Mouse Clicks Type Detection

    by Romy Budhi Widodo, Devina Trixie, Windra Swastika

    The operation of a computer required humans to use several parts of their body. However, there were some conditions where humans cannot operate computers correctly or in a normal position; examples of these conditions were accident victims and people with disabilities. Therefore, a system was needed to help make it easier for these people to operate the computer. This study developed a system that can classify click types using EMG sensors, the K-NN method, and the SVM method. EMG sensors helped take data in the form of signals from human muscle contractions which will later be classified into left-click and right-click. At the same time, it was useful for classifying these types of clicks for the K-NN and SVM methods. Data from EMG sensors were trained using the K-NN and SVM methods using 54 data sets in each class, namely left-click and right-click classes. The K-NN method was trained using k=3, 5, 7, 9, and 11. The SVM method used linear kernels, Radial Basis Function (RBF), polynomials, and sigmoids. After that, the accuracy values of the two methods will be compared. The study has successfully classified the types of clicks based on the input from the EMG sensor using the K-NN method with the highest accuracy results using k=3, which was 81.81%, and the SVM method using polynomial kernels which were 84.84%. The highest accuracy value was obtained by comparing the two methods, namely using the polynomial kernel SVM method. Adding datasets and conducting experiments using other methods as further comparisons can be used to improve system accuracy.

  • Open Access

    Perbandingan Particle Swarm Optimization dengan Genetic Algorithm dalam Feature Selection untuk Analisis Sentimen pada Permendikbudristek PPKS-LPT

    by Gregorius Guntur Sunardi Putra, Windra Swastika, Paulus Lucky Tirma Irawan

    Kekerasan seksual merupakan salah satu masalah serius yang perlu diselesaikan. Setiap tahunnya terdapat lebih dari 2000 aduan kekerasan seksual. Di satuan Pendidikan, sebanyak 55% dari pelaku merupakan guru atau tenaga pendidik. Jumlah yang ada masih belum merepresentasikan jumlah sebenarnya karena 63% dari peserta survei DIKTI tidak melaporkan kasus kekerasan seksual yang diketahuinya dengan alasan menjaga nama baik kampus. Dengan begitu, Kementerian Pendidikan, Kebudayaan, Riset, dan Teknologi (Kemendikbud) menerbitkan Peraturan Menteri Pendidikan, Kebudayaan, Riset, dan Teknologi (Permendikbudristek) Nomor 30 tahun 2021 tentang Pencegahan dan Penanganan Kekerasan Seksual di Lingkungan Perguruan Tinggi (PPKS-LPT). Peraturan ini menimbulkan kontroversi di masyarakat sehingga perlu dilakukan klasifikasi sentimen. Klasifikasi sentimen dilakukan menggunakan algoritma Multinomial Naïve Bayes, Complement Naïve Bayes, Support Vector Machine, dan Nearest Neighbor dengan menggunakan data seleksi fitur Genetic Algorithm dan Particle Swarm Optimization pada data hasil ekstraksi fitur TF-IDF dari data Twitter yang diambil pada tanggal 1 September-30 Desember 2021 menggunakan kata kunci “permendikbud”, hashtag “cabutpermendikbudristekno30”, dan hashtag “dukungpermendikbud30”. Kemudian klasifikasi akan dievaluasi menggunakan 10-Fold Cross Validation. Hasil dari penelitian ini, algoritma CNB memiliki kinerja paling baik jika dibandingkan dengan 4 algoritma klasifikasi lain dengan nilai accuracy 0.688, precison 0.685, recall 0.688, dan f1-score 0.683. Seleksi fitur GA belum mampu menemukan fitur-fitur yang memiliki akurasi lebih baik jika dibandingkan menggunakan semua fitur, sedangkan seleksi fitur PSO mampu menemukan fitur yang memiliki akurasi sama atau lebih baik jika dibandingkan menggunakan semua fitur dengan peningkatan tertinggi sebanyak 0.8% pada MNB. Sebanyak 49% dari responden merupakan pendukung pengesahan Permendikbudristek PPKS-LPT.


  • Open Access

    CNN Based Transfer Learning for Malaria Parasite Detection Using Thin-Blood Smear Images

    by Windra Swastika, Benedictus Julian Pradana, Romy Budhi Widodo, Rehmadanta Sitepu, Gregorius Guntur Sunardi Putra

    Transfer learning has been used in computer vision research, including in the health sector. In the health sector, the input image is generally an x-ray image or a microscopic image. In this study, transfer learning for models that have been trained using CNN to detect malaria parasites in red blood cell images. The deep CNN pre-trained model uses 3 architectures, namely ResNet50V2, EfficientNetB0, and InceptionV3. For each architecture, experiments will be carried out and compare which architecture is better in detecting malaria parasites. Based on experiments conducted without fine tune, the accuracy ranges from 0.76 – 0.81 for ResNet50v2, 0.76 – 0.80 for EfficientNetB0, and 0.77 – 0.82 for InceptionV3. The dataset is a collection of Blood Smear images which have two classes, uninfected and parasitized. The total number of datasets is 27,558, which is divided into two classes with the same number of different image sizes.


  • Open Access

    PREDIKSI HARGA FOREIGN EXCHANGE MATA UANG EUR/USD DAN GBP/USD MENGGUNAKAN LONG SHORT-TERM MEMORY

    by Angel Joanna Wijaya, Windra Swastika, Oesman Hendra Kelana

    Foreign exchange (Forex) adalah perdagangan pasangan mata uang dari harga mata uang suatu negara terhadap mata uang negara lainnya. Pada penelitian ini menggunakan metode Long Short-Term Memory (LSTM) untuk memprediksi harga close mata uang EUR/USD (Euro terhadap Dolar Amerika) dan GBP/USD (Pound Sterling terhadap Dolar Amerika) pada candle D1 (1 hari) dengan input harga open dan close. Hasil yang diperoleh model EUR/USD dengan 1 input mendapatkan nilai Mean Squared Error (MSE) terendah yaitu 0,0535 dengan model 1 layer LSTM 10 node dan menggunakan optimizer Nadam. Pada model 3 input mendapatkan nilai MSE terendah yaitu 0,0529 dengan model 1 layer LSTM 10 node dan menggunakan optimizer Nadam. Sedangkan, model 5 input mendapatkan nilai MSE terendah yaitu 0,0469 dengan model 1 layer LSTM 10 node dan menggunakan optimizer Nadam. Pada mata uang GBP/USD, model dengan 1 input mendapatkan nilai MSE terendah yaitu 0,0543 dengan model 1 layer LSTM 10 node dan menggunakan optimizer Nadam. Pada model 3 input mendapatkan nilai MSE terendah yaitu 0,0520 dengan model 1 layer LSTM 10 node dan menggunakan optimizer Adam. Sedangkan, model 5 input mendapatkan nilai MSE terendah yaitu 0,0631 dengan model 1 layer LSTM 10 node dan menggunakan optimizer Adam. Model dengan MSE terbaik yang digunakan pada website yang dibuat. Rekomendasi yang diberikan adalah pada mata uang EUR/USD menggunakan model 5 input dan mata uang GBP/USD menggunakan model 3 input. Hasil loss MSE, akurasi MSE, dan jumlah profit yang didapatkan semuanya adalah yang terbaik jika dibandingkan dengan model dengan jumlah input lainnya


  • Open Access

    Automatic Extraction of Diaphragm Motion and Respiratory Pattern from Time-sequential Thoracic MRI

    by Windra Swastika, Yoshitada Masuda, Takashi Ohnishi, Hideaki Haneishi

    Thoracic time-sequential MRI can be used to assess diaphragm motion pattern without exposing radiation to subject. Clinicians may employ the motion to evaluate the severeness of chronic obstructive pulmonary disease (COPD). This study proposed a novel method of diaphragm motion extraction method on time-sequential thoracic MRI in sagittal plane. Otsu’s threshold and active contour algorithm are used to obtain diaphragm boundary. An automatic diaphragm motion tracking and extraction of respiratory pattern are also performed based on the diaphragm boundary. A total of 1200 frames time-sequential MRI in sagittal plane was obtained for total of 15 subjects (8 healthy volunteers and 7 COPD patients). The proposed method successfully extracts diaphragm motion and respiratory patterns for both healthy volunteers and COPD patients.

  • Reduction of acquisition time in the intersection profile method for four-dimensional magnetic resonance imaging reconstruction of thoracoabdominal organs

    by Windra Swastika, Yoshitada Masuda, Takashi Ohnishi, Hideaki Haneishi

    We have previously proposed an intersection profile method for reconstructing four-dimensional (4-D) magnetic resonance imaging (MRI) consisting of one breathing cycle of the thoracoabdominal region. This method captures a set of temporal sequence images in a proper sagittal plane and sets of temporal sequence images in continuous coronal slices. The former set is used as a navigator slice and the latter sets are used as data slices. A 4-D MRI is reconstructed by synchronizing the respiratory pattern found in the navigator slice and the data slices. We propose a prospective method to reduce the acquisition time for data slices. During data slice acquisition, the synchronization process between the respiratory pattern found in the navigator slice and one data slice is monitored in real time. Data acquisition will be terminated and moved to the next data slice based on a threshold value. We used 14 data sets (seven patients with certain pulmonary disease and seven healthy volunteers) previously obtained for the original intersection profile method for a simulation using the proposed method to evaluate the time reduction and impact on image quality. Each of the data set was tested using three different threshold values and the acquisition time can be reduced up to 75%. Although the quantitative evaluation of image quality was slightly worse than that by the conventional method, the difference based on the visual inspection was subtle to human eyes.