3D ResNet Deep Learning Model Classifies Cisplatin Response in Colorectal Cancer Spheroids Using OCT Imaging

A deep learning model built on a 3D ResNet18 (搜索)搜索该靶点 architecture can classify the efficacy of cisplatin查看药品详情 in three-dimensional colorectal cancer (搜索)搜索该疾病 spheroid models imaged with optical coherence tomography (OCT), according to a study published in PLOS ONE. The work positions volumetric OCT imaging combined with convolutional neural networks (CNNs) as a potential route toward fully automated, high-throughput drug screening.

Spheroid Model and Drug Exposure

Heterospheroid cultures were prepared by co-culturing HT-29 (搜索)搜索该靶点 colon cancer cells with 1BR.3.G fibroblasts in ultra-low-attachment 96-well plates at a 1:2 ratio (2,500 cancer cells to 5,000 fibroblasts per well). Co-cultures were incubated for 72 hours to allow spheroids to form before intervention.

Cisplatin查看药品详情 dilutions were prepared from a dimethylformamide (DMF) stock solution (33.3 mM) in culture media, with a vehicle suspension containing 0.3% (v/v) DMF. At day 3 after seeding, each plate was treated with vehicle, 25 µM cisplatin, or 100 µM cisplatin, with 20 replicate wells per condition. After 4 days (96 hours) of drug exposure, spheroids were embedded in Histogel for OCT imaging, yielding 180 spheroids per treatment and 540 spheroids in total.

Validation of Treatment Effect

Bright field imaging with a NYONE Scientific Imager showed clear morphological differences between treatment groups. Vehicle-treated spheroids appeared intact and generally larger 96 hours after treatment, while spheroids exposed to 25 µM cisplatin查看药品详情 did not appear to grow, with evidence of degradation including diffuse haze and increased debris in the media. Spheroids treated with 100 µM cisplatin showed further degradation, including distorted non-spherical shapes and a larger hazy region of cell debris surrounding the spheroid.

Histological analysis confirmed these observations. Sections stained with hematoxylin and eosin and by immunohistochemistry for cytokeratin (搜索)搜索该靶点 (cancer cells) and MRC2 (搜索)搜索该靶点 (fibroblasts) showed that in vehicle-treated spheroids, fibroblasts assemble in the interior while cancer cells form the outer shell with a few cells infiltrating the core fibroblast region. Histology of spheroids treated with 100 µM cisplatin查看药品详情 could not be performed because of the extent of degradation and loss of structural compactness; paraffin embedding after fixation was unsuccessful due to disintegration.

Metabolic assessment using the CellTiter-Glo 3D Cell Viability Assay confirmed a significant decrease in metabolic activity between all three tested groups (p < 0.0001), based on 20 replicate samples per condition analyzed by one-way ANOVA with Dunnett's multiple comparison test.

Model Architecture and Performance

The researchers compared two CNN designs based on ResNet18 (搜索)搜索该靶点, selected to accommodate hardware limitations and ensure compatibility with consumer-grade systems. A 2D multi-view CNN processes four cross-sectional slices per spheroid volume—two orthogonal vertical slices and two vertical slices at 45 degrees to them—with output feature vectors concatenated and passed through a multi-layer perceptron. A 3D CNN processes a down-sampled version of the full 3D volume (200×500×500 voxels) on a GeForce RTX 4090 with 24 GB RAM.

Of the 540 spheroids imaged, 467 images were included in the dataset. Images were excluded when spheroids were destroyed during Histogel embedding, leading to empty wells, or when spheroids were improperly placed in the well—for example, at the surface of the Histogel rather than embedded at the bottom. These outlier images were removed because their inclusion significantly reduced classifier performance.

The confusion matrices showed that when either model predicted the wrong class, it selected the neighboring class; an image from the control group was never falsely predicted to belong to the 100 µM cisplatin查看药品详情 group. The authors note that the most accurate model, the 3D CNN, appears unbiased, as it is trained on all voxels in the image rather than relying on a user-defined set of 2D slices. These factors indicate that treatment causes regression patterns whose quantity changes linearly with dose—promising for building a model of structural change for arbitrary treatment doses, though the authors state this will require a large amount of additional image data across more treatment doses.

Attention Maps Reveal Dose-Dependent Focus

Attention maps generated for each spheroid showed that in the vehicle control group, the 3D model focused on the entire spheroid, identifying the full structure as relevant. With increasing cisplatin查看药品详情 concentrations, focus shifted toward debris outside the original spheroid. At 25 µM, the main spheroid structure still influenced the model, but increased cell debris at the bottom of the well emerged as a notable feature. At 100 µM, the main spheroid structure was no longer identified as a key feature; surrounding disintegrating cell material was a stronger identifying feature. There was no significant attention on the background.

Qualitative inspection across all samples revealed three recurring spatial patterns: spheroid-focused attention, plate/debris-focused attention, and a mixed distribution. In the control group, spheroid-focused attention dominated with no plate-focused cases observed. The 25 µM group introduced a substantial proportion of mixed-signal samples alongside a small number of plate-focused cases, while the 100 µM group showed the highest proportion of plate-focused cases and a considerable number of mixed cases, tracking directly with the degree of morphological disintegration. The authors conclude that the plate-contact region, where cell debris accumulates following cytotoxic treatment, represents a genuinely informative spatial feature rather than a simple imaging artefact.

Context and Limitations

The authors note that prior studies have demonstrated OCT for spheroid imaging, including more accurate assessment of spheroid volume, detection of necrotic core regions, and investigation of apoptosis and cell viability, as well as detection of drug-induced changes. However, most of these studies used only cross-sectional B-scans; the present work shows that volumetric data can improve classification and segmentation tasks and refine the set of key image features relevant for assessing drug efficacy.

Several limitations are acknowledged. The model was trained on images of the same spheroid type treated with the same drug, and given the sensitivity of such classification models to small variations in images, it is unclear how easily the model can be expanded to recognize responses to other drugs—a step the authors describe as critical for identifying novel drug candidates. Embedding spheroids in Histogel improved the signal-to-noise ratio, particularly by reducing reflection from the bottom of the well plate, but for spheroids treated with 100 µM cisplatin查看药品详情 most cell debris was removed during embedding, indicating some information loss. Histogel also scattered light, increasing background noise. The authors suggest that a 2D classifier could be incorporated to remove outliers automatically in advance, given the impact of outlier removal on performance.

The model was deployed using consumer-grade hardware without access to specialized computing resources such as supercomputers, demonstrating potential for wider impact, though the authors acknowledge that such resources could improve accuracy and training times. In a future study, the researchers intend to further investigate the dose-dependent response to confirm whether the model could be used to select optimal drug concentrations as part of the drug screening process, and to explore connections between specific drug mechanisms of action and their OCT image hallmarks.