A significant proportion of the cancer related deaths is attributed to drug resistance (>90%). This is a serious concern for cancer patients as resistance to the given therapy leads to reduced drug response and disease recurrence. A healthcare solution that is automated, cost-effective, and reliable, is the pressing need of this hour. However, realizing this goal requires studying and benchmarking computer vision tools and techniques for establishing a robust image biomarker for drug profiling. At present, the current state-of-the-art in visual recognition, i.e., the deep convolutional neural network, offers black-box solutions to image recognition problems. In contrast, the field of biomedical science often requires interpretability as a strong criterion for any expert system to aid in the diagnosis and therapy planning. Though the traditional machine learning techniques allow one to build engineered and easy-to-interpret features, those methods are found to provide inferior results when compared with the performance of deep learning. In this joint venture on AI and health, our objective is to obtain the best of both the worlds, leveraging the excellent pattern recognition power of deep convolutional neural networks while driving the decision-making process from an intuitive, black-box free perspective. Perhaps the best value that comes with the old school of machine learning is the opportunity to explore a dataset: manifold learning, cluster identifications, feature selections - these constitute an excellent set of tools that empowers a biomedical expert to develop core understanding of the datasets. To facilitate this process, we realize that building a large database associated with various annotations is key to our success. We are building a large databse of tumour spheroids, and characterizing them, based on different drug responsiveness to chemotherapeutic agents. The spheroids will be monitored in a real time imaging system. The effective in-vitro classifications are expected to provide the effectiveness of an existing or a new chemotherapeutic drug and the extent that the tumour can offer resistance to that particular drug. By capturing dominant patterns and hidden behaviour in image sets of tumour spheroids, the proposed computer vision solutions can provide a complete end-to-end apparatus for devising personalised therapeutic strategy while recommending the suitable alternatives if the first line of therapy fails.
We pose the high throughput, automated monitoring of spheroid images as a supervised. machine learning task with three classes of spheroid images, namely, Sensitive, Moderately Sensitive, and Resistant (Figure 1). The objective of the multiclass-AI system is to classify a novel test image into its most appropriate class .
Datasets and Python API
Pre-treated Post-diagnosed (PTPD) HNSCC cancer cell lines are used to develop
the spheroids. We have a plan to establish at least 10 new HNSCC spheroids.
In brief, the HNSCC cell lines (monolayer or 2D) will be maintained in complete
keratinocyte serum-free growth medium. For the spheroid initiation,
200 µL of single-cell suspensions will be seeded in ultra-low attachment
(ULA) plates at varying cell densities. The plates will be incubated at a
humidified 5% CO2 atmosphere at 37ºC (48-72 hrs) for maturation.
Progression of spheroid formation is imaged on a daily basis (24, 48, and 72 h) using an optical microscope with 5× or 10× objectives. The formation of tumour spheroids will also be monitored every 3 hours by live-cell imaging using Incucyte Zoom™ throughout the entire spheroid formation process with a phase-contrast/fluorescence set up using the 10× objective. The images are annotated according to drug sensitivity tests and then passed on to the computer vision modules for Image analysis.
The image dataset is curated, annotated, and equipped with Python API. We have undertaken a benchmarked study with the existing computer vision tools and techniques. The results are very promising. Soon we shall release the entire tumor spheroid dataset, annotation, and the codebase for scholarly pursuits.