msf-CNN efficiently uses patch-based fusion to optimize CNN inference execution on MCUs, reducing RAM usage by up to 50% compared to prior methods while offering greater design flexibility.