RNN vs Dense questions for two dimensional data

I’m making a network to predict an single number based on a 2d graph. The input would be have dimensions ~2691x2 (or 2x2691? not sure which is convention) but each “graph” will have a bit more or less than 2691. Around that number, but not exactly. Because the dimension changes, would it be best to run the each point on the graph through an LSTM and predict ? Would I use the hidden state of the LSTM or the final output of the LSTM? My gut tells me hidden state, as it represents the content of the entire graph and output is just the next output. Otherwise, I would do this with an MLP and pad up to a maximum number (say 3000 or so), or a CNN and convolve with a sliding window type thing.