Enterprise AI Analysis

Solution Proposed: An explainable deep learning framework is proposed for reliable species-level classification of three Chlorophyta microalgae (Desmodesmus flavescens, Desmodesmus subspicatus, and Tetradesmus dimorphus) using bright-field microscope images. Twelve CNN architectures (EfficientNet B0-B7, DenseNet201, NASNetLarge, Xception, ResNet152V2) were benchmarked with standardized preprocessing and data augmentation. Interpretability was ensured using Grad-CAM and saliency maps to visualize model decision-making.

Methodology: A dataset of 3624 microscopic images was curated, split 70:15:15 for training, validation, and testing. Images were cropped to isolate organisms, resized to 224x224, converted to grayscale, and contrast-enhanced using CLAHE. Offline data augmentation (flipping, rotation, zooming, brightness) was applied to the training set. Models were initialized with ImageNet weights, trained with Adam optimizer (LR=1e-4, batch=32) for 50 epochs with early stopping, and evaluated using accuracy, precision, recall, F1-score, and ROC/AUC metrics. XAI techniques (Grad-CAM, saliency maps) were applied post-training for interpretability.

Results & Findings: ResNet152V2 achieved the highest overall performance with a macro F1-score of 0.975, outperforming other models. Most models showed strong performance with macro F1-scores generally exceeding 0.90. Learning curves indicated consistent convergence without overfitting. The confusion matrix for ResNet152V2 showed near-perfect classification for Tetradesmus dimorphus (100%) and Desmodesmus flavescens (98.77%), with D. subspicatus at 93.67%. XAI visualizations consistently highlighted biologically relevant features like cell walls, surface ornamentation, and colony structures, confirming models used taxonomically meaningful cues. AUC values were consistently near 1.0, indicating robust discriminative abilities.