What can i do to make train loss, accuracy and validation loss, accuracy normal?

hello,it is my first time to ask a question in pytorch forum!
when i train a 3dunet in cremi dataset to segment synapses, lower accuracy is better (update now, best score is about 50). However, my results are very bad just like the picture below.
i want know what reason will result in so bad result ?
and how can i solve ?
thank you !