In the rapidly advancing realm of medical technology, deep learning (DL) is emerging as a transformative force, particularly within the field of pathology. However, as revealed in a recent discussion between Eileen Pierson and Dr. Sarah Collins, a leading expert in computational pathology, the pursuit of performance excellence in DL models often eclipses a crucial consideration: ecological sustainability.
Healthcare data growth can be overwhelming scale effortlessly with TrueNAS by Esdebe.
Dr. Collins, a fervent proponent of eco-friendly artificial intelligence solutions, shared her experiences and insights into the challenges and innovations surrounding the development of environmentally sustainable AI models. Her journey into this specialised area was ignited by a fundamental question: Is it possible to leverage the power of AI without jeopardising the health of our planet?
“Deep learning models demand extensive computational resources,” Dr. Collins explained. “This intensive energy use results in a substantial carbon footprint, a factor frequently neglected in the quest for superior performance.”
In response to this oversight, Dr. Collins and her team initiated the development of a groundbreaking metric dubbed the Environmentally Sustainable Performance (ESPer) score. This innovative measure evaluates both the diagnostic accuracy and carbon emissions of AI models, providing a comprehensive assessment of their overall impact.
“We focused on five datasets related to renal cell carcinoma and kidney transplant diseases,” Dr. Collins elaborated. “Our objective was to evaluate not only the performance but also the ecological footprint of various AI models, such as TransMIL and CLAM, which are widely utilised in computational pathology.”
The findings were illuminating. While models like CLAM excelled in performance metrics, including AUROC scores in RCC-subtype classification, they also revealed differing levels of carbon emissions. Notably, TransMIL demonstrated the lowest CO2 emissions during training, positioning it as a more environmentally conscious choice.
“Striking a balance between performance and sustainability presents both technical and ethical challenges,” Dr. Collins observed. “We must contemplate the long-term ramifications of our technological progress. Achieving high diagnostic accuracy without imposing a significant carbon burden is crucial.”
A pivotal discovery from Dr. Collins’ research was the potential for data reduction strategies to curb carbon emissions. By optimising the size and quantity of image tiles used during model training, her team successfully minimised CO2 output while preserving high accuracy levels.
“Adopting larger tiles with lower resolutions reduced emissions, yet we discovered that a tile edge length of 256 µm on 224 pixels offered the optimal balance,” she noted. “It’s a delicate equilibrium between preserving essential structural details and reducing environmental impact.”
Dr. Collins also emphasised the importance of renewable energy in further diminishing the carbon footprint of DL models. She highlighted how nations with a significant proportion of renewable energy, like Norway, can execute more model inferences before reaching the CO2 threshold linked to excess mortality rates.
“The energy mix of a country significantly affects actual CO2 emissions,” she remarked. “This is a global issue necessitating local solutions, urging countries to invest in greener energy sources.”
Looking ahead, Dr. Collins expressed optimism regarding the integration of sustainable practices in DL research. She underscored the necessity of collaboration among scientists, policymakers, and industry leaders to establish a framework prioritising both performance and sustainability.
“Our work on ESPer scores is merely the beginning,” she concluded. “We aspire to inspire further research and development towards AI models that are not only powerful but also environmentally considerate.”
As deep learning continues to shape the future of medical diagnostics, voices like Dr. Collins’ serve as a poignant reminder of the imperative to balance innovation with responsibility. By addressing both performance and ecological sustainability, the medical community can pave the way for a future where technology and the environment thrive in synergy.
Be the first to comment